/hdg/ - Stable Diffusion

Anime girls generated with AI

Index Catalog Archive Bottom Refresh
Name
Options
Subject
Message

Max message length: 12000

files

Max file size: 10.00 MB

Total max file size: 50.00 MB

Max files: 4

Supported file types: GIF, JPG, PNG, WebM, OGG, and more

E-mail
Password

(used to delete files and posts)

Misc

Remember to follow the Rules

The backup domains are located at 8chan.se and 8chan.cc. TOR access can be found here, or you can access the TOR portal from the clearnet at Redchannit 3.0.

US Election Thread

8chan.moe is a hobby project with no affiliation whatsoever to the administration of any other "8chan" site, past or present.

/hdg/ #12 Anonymous 04/04/2023 (Tue) 17:34:49 No. 13776
Previous thread >>12556
DId anyone open an issue with Kohakublueleaf's repo regarding the new updates that Kohya did with block weights?
>>13735 Artist LoRa is this guy: https://www.pixiv.net/en/users/66733693 I haven't released it anywhere cause it's kinda finicky, might at some point though. Does do meaty lolis quite well though.
So it looks like Civitai just issued some new woke policy and is starting to nuke content off their site. This trainwreck just keeps getting better and better.
>>13779 Still using it for numbers peko, what exactly did they do?
>>13780 Some sort of Anti-Hate speech policy. The main thing being talked about was that a Prophet Mohammad LoRA was uploaded that potentially started the shit show as it is one of the ones deleted. https://civitai.com/models/27592/mohammad-the-prophet
hey all, Just finished my testing for the new resize and extract queue stuff. It all works and my readme has an explanation of it. Basically, I tried to make it really easy to do multiple resizes or extracts. I set up two ways to do it, through a txt file, which just contains a list of values, one per line, or putting them in directly in the popups. they work pretty well, and are a great way to resize larger dim lora a bunch of times and test them all out. other than that, updated to the newest LyCORIS, and the version right before the newest sd-scripts. you can update through the update bat file, make sure you also run upgrade.bat thanks everybody for using my scripts, knowing they get used keeps me motivated to continue working on them. I'll get the new sd-scripts update out asap, just got some logistical concerns to worry about. links: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts installer if you don't have them already: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/releases/tag/installers-v6
>>13761 Yeah this is too autismo, lora does everything I want good enough.
>>13783 I see use in it, but the average user really doesn't need it. I'm planning on at least fucking with it to see how it works and possibly ways to simplify it a bit, such as creating general presets that work most of the time, or whatever. I personally think it'd be cool to be able to just train the face, or just train the lighting, but those are niche so it's whatever
>>13782 Thanks anon, hopefully Kohakublueleaf has noticed Kohya's recent updates to implement their own method of block training with LOCONs and LOHAs but otherwise I'll be looking forward to the next major update as well
>>13778 know that feeling certain loras are not perfect which makes me hesistant to share them Don't want this place to end like civitai where one out of a hundred lora are useable
>>13786 I feel the same way, if I'm gonna share anything, I'm only gonna share it here with you guys.
>>13788 >penis of the lake Araragi sure can hold his breath
>>13786 >Release loras regularly for a bit >New tech comes, remake, reupload >new tech comes, remake >new tech comes, remake A lot of these at this point are on like v8-v13 from data set adjustments or just new tech from after they were made. It's just tiresome to reupload.
(1.62 MB 1024x1536 catbox_xwhxkv.png)

>>13788 Magnificent! Legends say the girl who gets inseminated by this magical creature shall bear the king and help him reign all of the kingdom... >>13790 perfection is a curse...
>>13790 the worst part of all this is the more loras you make the more of a perfectionist you become
>>13792 >>13791 And on top of the escalating perfectionism is I'm keeping multiple epochs of each since they do things better or worse and merge-ing and/or lowering dim doesn't address that or really do much. It's pain.
>>13783 Is that a Monika Lora?
>>13794 Yeah just finished baking it last night so hadn't fiddled with it much. It's probably gonna be re-done though. I took a gamble and added a whole doujin or two of her from KTKM or whoever that dude who loves Monika. Haven't played with it much but it becomes apparently quickly
(5.92 MB 1536x2304 grid-0034.png)

>training a locon for the first time while reading what could be done better >type something I don't remember >get these pics not even close what it should be how the fuck are these 9 different styles?
>>13796 Are you using the method that was implemented today?
>>13797 I updated easy scripts today so... I don't know?
(3.01 MB 1280x1920 catbox_gs8sgp.png)

What are frogs
>>13797 wow... now I feel dumb I changed the model to something else but based64... welp m ybad
>>13798 oh ok, that is still using the old method then, the new method where you can specify the dim for each block or was it learning rate, is going to be implemented later. I'm keeping both the old method and new method just in case because the new one is super autistic desu
(2.54 MB 1280x1920 catbox_mhs3et.png)

(2.53 MB 1920x1280 catbox_nn20gt.png)

(2.46 MB 1280x1920 catbox_oesa4d.png)

(2.49 MB 1280x1920 catbox_0wd8gy.png)

(6.95 MB 2304x2304 grid-0039.png)

>>13801 I am just an idiot who is unable to use to correct model and wonders why the output is strange but thanks!
>>13799 holy froggers anon, are you gonna share that lily LoRA?
(2.94 MB 1280x1920 catbox_cdqea6.png)

(2.40 MB 1280x1920 catbox_2xbdjz.png)

Here you go just wanted to test before uploading https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/7F8GTbyD
>>13805 I, for one, welcome our new froggy overlords Thank you, anon!
>>13805 based, she's been on my backlog but I don't have a lot of experience with small dataset wizardry I will do unspeakable things to my little sister when I get home later
(1.84 MB 1024x1536 catbox_ifltqb.png)

>>13807 look her in the eyes and say that again you wouldn't defile her, right? ... right?
>>13807 >small dataset what do you consider to be small dataset?
>>13809 a small dataset of 1 million images
>>13782 did you fuck up import requests in the installer? see https://boards.4chan.org/h/thread/7299074#p7299454
>>13809 Probably less than 30, and lily has around 50 pieces of fanart but most of it isn't her main outfit. >>13808 It was just a joke, I swear! I'm going to generate completely pure and innocent things with my sister!
(1.71 MB 1024x1536 catbox_v6dldf.png)

>>13812 https://gelbooru.com/index.php?page=post&s=list&tags=lily_hopkins+ our little sister has more than 50 pics of fanart never underestimate the urge to defile something pure
>>13813 well shit I guess I never checked gelbooru with the content filter off
>>13712 what is the hisha lora?
(2.41 MB 1920x1280 catbox_xiznbk.png)

(3.61 MB 1920x1280 catbox_jxkfr0.png)

>>13812 >>13813 Close, I actually ended up with just over 50 images total after pruning. Though a lot were from Girutea which I had set to repeat less so they had less impact on style
Lora merge gods when you want to retain the best of both loras do you do them at like .8 or .9 or just 1:1 on both and hope it works? I don't know shit about balancing these.
>every extension broken on my commit of the webui before the gradio update anyone remember how to make the current commit of the webui revert to a past one?
>>13818 first use "git log" to print out a list of past commits. Find the one you like then copy the string of random numbers and letters(hash) associated with it. Then use "git checkout <hash>"
I know we hate youtube posting here but I think I found the first youtube AIfag to actually post a not shitty guide to making LORAs https://www.youtube.com/watch?v=7m522D01mh0
>>13820 >did lycoris guide lmao that thing is pretty pointless now after today's update until KBL updates their repo
(1.13 MB 960x1152 catbox_dxdjcn.png)

(1.41 MB 960x1152 catbox_f87gd0.png)

this dress was built for butt shots
>>13820 >re-naming the image files and the dude lost me there, and you only really need to crop if you're being super autistic but alas my stupid past self never saved my old datasets that weren't cropped that would have been useful and less time wasting with the way I bake my LORAs now oof
>No updates for 2 day on the dataset extension Fuckers, you can't just leave it rotting like this, it's unusable
I don't have it set up yet, would this work well with controlnet?
>>13826 to clarify: I wanted to get this in color but all the coloring AIs I tried suck ass so I thought I'd do img2img instead with controlnet
Easy script anon, I got this error when installing the scripts using your installer. Funny thing is though, this happened only during a second install after I screwed something up and decided to delete everything and reinstall Python, while the first one installed just fine.
>>13826 if you bitcrush it to a 1 bit png it will work with canny yeah, just make sure to get rid of the shading first
(323.32 KB 748x1057 43745796_p0.jpg)

my dataset is pretty much entirely censored and I don't think the artist ever released the original uncensored drawings, what do? Is it gonna cause issues? i could find another artist who uses copics almost exclusively but this one is my favorite
>>13829 >just make sure to get rid of the shading first So just grab the bucket tool and color all the gray parts white?
>>13825 There really needs to be a minimalist UI for these extensions >>13830 based momiji enjoyer have to do shishi juuroku sometome soon
>>13832 >shishi juuroku based pudgy momiji enjoyer
>>13831 yeah, should work and presumably you're using an anime model which means it should infer clownpiece's colors just fine
some testing of negative embeds
>>13834 style is gonna depend on the model, right? i wanted to retain the original style but i guess that's too complicated for now
(8.83 MB 4864x3421 tmpmyo3a4sd.jpg)

(8.63 MB 4864x3419 tmpxuuh9yci.jpg)

>>13835 fuck sake files too large
>>13837 can I get a link to the bad image v2 embed?
>>13837 That's already so cursed I can't tell what's supposed to be better or worse.
>>13836 canny will preserve the shapes of the linework so you'll have forms in the style of the artist but not color (unless you train a lora on the artist to learn their coloring technique kek)
teto is a real commercial soundbank now so i'm gonna celebrate by making way too much porn of her
>>13840 the base negative is just (chibi:1.3) so take that as a cautionary lesson regarding western-ish loras and the lack of (low quality, worst quality:1.4) i do like the combination of easynegative and bad image 2 even if it really homogenizes the face
Finished reworking the tags on the Seraziel dataset (along with adding the latest images). Still not going for a quality separation for repeats, for now I only separated for the cock type. Only like 10-15% bloat from image dupes between these folders, not a big deal especially since they're mostly quality ones. I'll tackle the cleaning tomorrow (about 1k images lmao)
(3.54 MB 1433x2150 catbox_ynf1su.png)

(3.36 MB 1433x2150 catbox_rtxz6i.png)

(3.34 MB 1433x2150 catbox_ewevqm.png)

>>13847 who is this hot whore
>>13811 the installer should install the requests library on first run if it isn't installed, so that shouldn't be an issue. I tested the installer again before I released the update and it worked fine for me, let me to some testing, I'll look into it.
>>13849 yeah, installer works fine for me, must be how the launched it. might have to look into making a proper exe installer. so people stop launching the installer incorrectly
(1.06 MB 896x1024 grid-0014.png)

(1.18 MB 1024x896 grid-0009.png)

Anyone have any tips for generating shota? My only success is using a style lora for shota artists but when I try to add a character lora it gets too cooked
>>13816 Love the second image.
>>13850 hmm, I tried it again on a different computer and it gave me the error... hmm, I think I know what happened, I might be able to fix it
(3.24 MB 1920x1280 catbox_anm3h0.png)

(3.23 MB 1920x1280 catbox_56xk88.png)

(2.97 MB 1280x1920 catbox_x0fa7r.png)

(3.06 MB 1280x1920 catbox_47ragj.png)

A few more
>>13853 yep I figured it out, I was an idiot and forgot to re import requests after it was installed. that being said, gonna look into changing it into an exe so that people can just double click to install instead of having to run it through the command line
>>13854 >the feet aren't mangled teach me your secrets
(3.30 MB 1280x1920 catbox_onpg11.png)

(6.02 MB 5120x1187 catbox_nfh7h5.png)

>>13856 Just a good combination of loras and a little gacha https://files.catbox.moe/stvk9j.png
(591.40 KB 3456x1468 catbox_jyv5f6.jpg)

thank you lilyhops anon I was looking at some of your catboxes and realized I've been fucking retarded using latent upscaler for everything
So apparently my old gradio commit no longer wants to work with the auto tag extension... but the latest commit with the gradio update does, man what the hell
(540.43 KB 512x768 catbox_9om6qn.png)

>>13859 Yamamoto souichirou lora? That was the first embed I tried to make before loras. Needless to say it didn't go very well
>>13830 If you tag with "censored" as appropiate then you should be able to put it in the negatives and "uncensored" in the prompt. There's been several LoRAs trained on hentai screencaps that take advantage of this
so, this style lora kinda lolifies everything. kinda based tbh >>13864 yeah
To be honest I would have no issue switching webuis but every time I had bullshit issues like extensions that used to work no longer working anymore I just decide to do a fresh install but moving everything is such a pain in the ass, this 2tb gen 3 NVME is already weird as shit because there's the issue where I'm mixing models and if I open something that takes up as much file processing as data mixing it'll just make the ssd go *poof* and I literally have to reinstall the NVME just to get it recognized by windows again, that has nothing to do with the webui just me ranting about how stupid things can get when they don't work
>>13863 did you try deleting the auto tag and reinstalling on the old commit?
>>13867 haha slc cache goes BRRRRRRRRRRRRRRRRRRRRRRRRRRR
>>13862 people kept telling me to use remacri but it's not THAT good, what are your settings for 4x-animesharp?
(2.78 MB 1280x1920 catbox_084dwa.png)

>>13862 No problem. Historically I've been using 4x-Ultrasharp with 0.4 denoise and only a few hires steps for quick gens, especially for under 1.5x upscale. I've only been trying out other upscalers since I've been able to upscale 2x from 640x960 without getting cuda memory errors. It takes longer but the quality is worth it.
>>13868 Yeah I did, it just gives me a python error that I can't pinpoint, I can reinstall it again and post the error here just give me a sec
So wait, Lollypop is the definitive Hires upscaler?
>>13870 the meatadata is in the image, it's 1.5x, 10 steps, 0.55 denoise for all of them
>>13869 It's literally when I open genshin impact and have it do the file integrity shit and when I'm mixing models that's when my poor NVME just ends up disappearing, literally no one has an answer for that shit on google I just have to pop it out and pop it back in to get to work again... >good health my ass
>>13874 x/y plots keep the metadata? huh
ok so first sd-scripts updates to add a bunch of block weight shit, and now LyCORIS updates to add a new module called (IA)^3 god dammit, now I need to release an update adding that before I stop supporting LyCORIS for the time being so people can try this out, if you want. Seems like another LoHa type thing based on the comments in the code but I haven't looked into it
>>13875 which drive is it and for the love of god don't tell me a crucial p5 i just bought three of them sounds like some super low end nvme whose cache gets fucked when you process too much shit at once, old intel 660ps do that
>>13877 I gave up on LOHA, it was only LOCON I was interested in, unless KBL implements the block weight shit that Kohya did I'll be interested >>13878 Lexar...
>>13879 yeah, it's more of an exasperated "fuck, now I gotta support this" than a "Oh cool a new update from LyCORIS"
>>13880 I never got the hang of LOHA, after all the baking I did I just decided to stick with LOCON which never ended up using block weights correctly but the NEW method of using it is literally so autistic that I have to get some headache meds ready for that because it's going to be worse than all those countless tests I did to try to my my LOHAs look as good as my LORAs/Locons
>>13861 How did you get the sleeves? My model always puts hands on the end of the sleeves.
>>13881 techincally with the block weight way, you can just specify all the weight to be 1 and it'd be close to the default... I think? I honestly haven't looked into it enough, either way, gonna get the update out for the new LyCORIS update so I have at least one update that supports it.
Alrighty pythonbros, here's the error I get with the commit I have before the gradio update when I try to use the auto tag complete option File "M:\AI\Auto111 Old Gradio\stable-diffusion-webui\venv\lib\site-packages\gradio\routes.py", line 337, in run_predict output = await app.get_blocks().process_api( File "M:\AI\Auto111 Old Gradio\stable-diffusion-webui\venv\lib\site-packages\gradio\blocks.py", line 1018, in process_api data = self.postprocess_data(fn_index, result["prediction"], state) File "M:\AI\Auto111 Old Gradio\stable-diffusion-webui\venv\lib\site-packages\gradio\blocks.py", line 947, in postprocess_data prediction_value = postprocess_update_dict( File "M:\AI\Auto111 Old Gradio\stable-diffusion-webui\venv\lib\site-packages\gradio\blocks.py", line 371, in postprocess_update_dict update_dict = block.get_specific_update(update_dict) File "M:\AI\Auto111 Old Gradio\stable-diffusion-webui\venv\lib\site-packages\gradio\blocks.py", line 257, in get_specific_update specific_update = cls.update(**generic_update) TypeError: Textbox.update() got an unexpected keyword argument 'multiline' never got this error with the latest commit
>>13884 even tried installing the auto tag extension through using the code on the github page itself but nope still get that dumbass error, it's literally the only extension I install that gives me this weird shit on the commit before the major gradio update
(1.82 MB 960x1440 catbox_7rzq5n.png)

(1.67 MB 960x1440 catbox_en7wzo.png)

(1.72 MB 1152x1152 catbox_1dmosp.png)

>>13876 yup, even the jpg downsized outputs have it
>>13879 >Lexar... the NM series is unbelievably bad, if you want a cheap decent nvmes get a crucial p3 plus (not sponsored but i just bought way too many fucking crucial drives in the last year and none of them have failed me unlike samshit's)
>>13887 nah no worries I plan on buying a bunch of NVME SSDs to make use of the pcie expansion card I got with this motherboard, I'll keep them in mind I just need more space for LORA training and model saving
>>13888 if you're on X670/E just remember that that kind of card will probably make your x16 gpu slot drop to x8 not a problem yet since pcie 5.0 gpus aren't a thing yet but y'know
>>13890 well it will be ironic if the 5000 series ends up actually making use of a PCIE 5.0 slot but we're still 2 years from that or less now Jewvidia
>>13862 latent hiresfix is great when everything's copacetic but lollypop is good at paving over minor lora/model disagreements. so's euler a, for that matter. >>13889 insanely based, thank you
>>13891 >12vhpwr >12VO >no more 24-pin MB connectors >ATX 3.0 >PCIE 5 pc building got infinitely gayer in less than a year
>>13893 I'm still looking for a 1600 watt PSU that doesn't cost more than my motherboard kek
>>13894 >that doesn't cost more than my motherboard kek gave up and settled for a 1300W one with a 12VHPWR connector for muh future-proofing 1650W GF3 is a bust with multiple people reporting crashes and shutdowns and something like anus' 1600W THOR is like.. ~900€. absolutely fucking insane (they don't even have a 12vhpwr connector) evga doesn't make any atx 3.0 ones (yet) and from what i know the 1600W ones have been good but not great
To anyone trying the recent dev builds for 2.1.0 torch on the webui, you'll get a million errors about cuda and facebook employees if the xformers you are using was built on a different 2.1.0 build date than the one you installed with the webui
>>13896 NotImplementedError: Could not run 'xformers::efficient_attention_forward_cutlass' with arguments from the 'CUDA' backend. This could be because the operator doesn't exist for this backend, or was omitted during the selective/custom build process (if using custom build). If you are a Facebook employee using PyTorch on mobile, please visit https://fburl.com/ptmfixes for possible resolutions not sure if that means the current build I made on the 3rd would work with older dev builds, I'm too lazy to test it myself honestly
>>13897 I'm going to test it out anyways, it's going to be BS If xformers build on a later devbuild date doesn't end up working with an older devbuild >building xformers everytime there's a new devbuild for 2.1.0 now I understand why people call it "experimental"
4chin /hdg/ is reminding me of /vt/ with how angry and schizo some anons are now kek
>>13900 yeah that shit got real toxic (and I hate using that word)
>>13900 don't want to visit, what's the newest topic there? last time we were in the "/g/ refugees posting granny porn and/or anything but porn and c-model users calling regular /h/ material /aco/ material" week
>>13898 OK yeah it doesn't work, building xformers is such a pain it literally is a million of commands repeating themselves for 20 minutes until it's done building just wait until 2.1.0 is done and just use 2.0.0 torch with the webui
>>13902 Just talking about >MUH KEEP TOKENS DOESN'T MATTER >THIS PAPER FROM 2016 SAYS THAT BATCH 2 IS BEST >YOU FAGGOT just the basic stuff and schizo posting that melts my brain, but yeah my brain is hurting enough from shit not working with the webui now
>>13904 oh my fucking god we're still stuck in november
>>13903 no I'm retarded it does work, ok so xformers built on a later build than earlier devbuild for 2.1.0 will work, how did I fix it? Just delete the xformers folder I had before and reinstall it with the one I built later on. Yeah so basically >xformers built later will work on older builds >xformers needs to be made with the same 2.1.0 build date as the latest 2.1.0 build if you want to utilize it without the dumbass facebook employee error
>>13905 >oh my fucking god we're still stuck in november loras weren't a thing in november. and if you're talking about embeddings, i don't remember people arguing over batch size for that
(2.01 MB 1024x1536 00077-4077253222.png)

>>13900 yeah it's partially my fault for arguing with the schizo but it just pisses me off seeing people asking questions about training and watching him feed that shit to them
>>13896 this is why you just manually build xformers. takes like 3 commands and you only have to do it once
>>13909 >>13906 you only have to build if you're using a latest build, otherwise any xformers I made yesterday will work with all previous builds
>>13900 /vtai/ is alright, but yea /vt/ has left me feeling mentally drained a lot of times, especially when that one schizo goes on a rampage because someone else cracks a joke about their anime girl probably having a boyfriend or is fucking their dog and you got to somehow deal with the collateral of autistic posts flying around.
>>13790 >ishmael god I need this
>>13911 I'm that schizo, I want my oshi to never be near a man on stream, inshallah peko
>>13913 Ahhhhhhhhhhhhhhhhhhhhhh
More esrgan model tests to clean up datasets. 1x_DeEdge_105000_G seems to be best for removing halos on shitty oversharpened downscales.
(1.43 MB 1024x1536 catbox_7k2sri.png)

i'm really enjoying this lora
>>13916 that is great artist to make a lora of. can't believe that didn't cross my mind
>>13916 5heads... 5heads for days...
>>13917 if you're looking for ideas, i'm pretty sure nobody did one of the dagashi kashi guy yet. cot_510 on twitter.
>>13919 >kotoyama thanks for the idea
>>13915 Original -> 1x_JPEG_80-100 -> 1x_DeEdge_105000_G Always de-jpeg first. Always de-jpeg any jpeg
>>13921 is this a char or artist lora?
(2.14 MB 3840x4096 1680615978607670.jpg)

Reposting, anyone know the artist?
>>13895 still running my 4090 off of 1000 watts lol. imo not worth to get a powerhouse psu until the next gpu upgrade i make, which by then those psu's should be cheaper
>>13782 >>13828 I should have probably reply so you can see it. Better late than never I guess.
>>13926 bro your eyes? >>13853 >>13855
>>13915 >>13922 Any tips on running these on a batch of images?
(2.03 MB 1024x1536 catbox_9mwzgt.png)

(1.92 MB 1024x1536 catbox_hw1p9s.png)

(1.89 MB 1024x1536 catbox_ixbvwp.png)

(2.06 MB 1024x1536 catbox_v93vmd.png)

Forever river prompting.
>>13928 >one image outputted as a jpg when all others are pngs cursed shit
>>13928 Yeah I've been sing chaiNNer for batches so it's dead simple to set up and automate. >>13930 The jpg is the original image which is the whole point
>>13931 >The jpg is the original image which is the whole point No I'm converting jpgs to pngs with the 1x_JPEG_80-100 one. Now there are three images that outputted as jpgs and about 85 as pngs. All were jpgs before.
(2.19 MB 1024x1536 catbox_1jqxc0.png)

(1.91 MB 1024x1536 catbox_28o8q9.png)

(2.22 MB 1024x1536 catbox_ot7gyw.png)

>>13929 classics are classics for a reason the age old question is how much spice can you add without creating something entirely different?
(1.76 MB 1024x1536 catbox_b4iycm.png)

btw a serious question how smug is too smug? Take this for example: What does this expression tries to convey?
>>13931 >chaiNNer Do you use pytorch, ONNX or NCNN? Idk which one is the fastest
(1.81 MB 1024x1536 catbox_h2v72h.png)

(1.82 MB 1024x1536 catbox_k68pv7.png)

(1.78 MB 1024x1536 catbox_d3zqm1.png)

>>13933 Cute. >how much spice can you add without creating something entirely different? If you add so much spice that it somehow changes the setting of the prompt, it just has to be 1 girl or more in a river but I do wonder if you could make them ride a canoe or a raft.
(2.20 MB 1024x1536 catbox_dyl6gp.png)

(2.20 MB 1024x1536 catbox_wdmjr8.png)

(2.08 MB 1024x1536 catbox_uw4ogj.png)

>>13936 wouldn't that be exactly changing the setting from bathing to something else?
>>13937 Yes but it would be interesting if it actually works.
(2.02 MB 1024x1536 catbox_sp2816.png)

>>13938 everything is possible you just have to try and maybe it can compare to your imagination
>>13927 I didn't see a reply to my post so I thought he missed it. Didn't know other people reported the same thing too.
(1.86 MB 1024x1536 catbox_6upw3i.png)

(1.85 MB 1024x1536 catbox_op9cxw.png)

>>13940 The river prompters just keep on winning.
(2.20 MB 1024x1536 catbox_6gmu09.png)

>>13942 another example with highspeed canoe driving
>>13943 Now this is podracing
(1.70 MB 1024x1536 catbox_mf60ei.png)

(1.61 MB 1024x1536 catbox_1ff6o7.png)

(1.61 MB 1024x1536 catbox_mv6myg.png)

(1.54 MB 1024x1536 catbox_uixmza.png)

>>13941 sorry, yeah, I already noticed and fixed it, but I had to sleep, I'll update the installer in a small bit
that opossummachine lora is really good at lower weights, too, consistently helps a lot with hands and feet
Seraziel lora maker here, should I get rid of the usual rectangles he puts on images? I don't find them particularly bad on most (generated) images, see >>13526 Maybe I'll tag them? Like "sqare frame"
>>13941 sorry it took so long, I ended up being busy most of the morning, but I updated the installer to not break anymore https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/releases/tag/installers-v6
>>13951 lama cleaner, it will be very easy to remove. https://github.com/Sanster/lama-cleaner
>>13946 Catbox on the last one?
>>13953 though, you'll have to rebake it if you already had baked it
>>13935 Pytorch, there are some models that have onnx or ncnn variants (I think Ultrasharp or Animesharp) but I already had a bunch of models already downloaded they're more widespread. >>13932 Weird, I haven't run into anything like that. Here's my chaiNNer setup for reference
>>13953 >>13955 Yeah that's already what I'm using to remove signatures. I'll just tag them afterwards, it's not annoying like watermarks from my tests (appears rarely). >>13956 Yeah, I forgot that .pth was pytorch format so obviously I had no choice. The jpg to jpg bug was on webui, had no problem with chaiNNer, it's just very slow, took like 5-6 hours for 1k images total.
(1.40 MB 864x1304 06637-16605276.png)

(1.47 MB 864x1304 06612.png)

Can you recognize this artist anon? :^)
hey all, just did a small update to push the new ia3 module in LyCORIS, gonna start work on the big update now, have fun with this... if you want, I'm probably not touching it. anyways links: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts installer: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/releases/tag/installers-v6
(53.57 KB 512x512 1f62d.png)

>>13934 "needs bratty child correction" is what it conveys
>>13958 Finished cleaning the 120 holo images Now on the the 880 others... AI that detects text and then pipes it to lama when?
(1.73 MB 1024x1536 catbox_sv2mjz.png)

(1.43 MB 1024x1536 catbox_aihgd0.png)

(1.88 MB 1024x1536 catbox_fz6f0i.png)

>>13961 You wouldn't defile such a nice young lady, right? You mean praying to god with correction... right?
>>13963 her forehead is the canvas and my cock is the paintbrush
(1.67 MB 1024x1536 catbox_tygjrt.png)

>>13964 Rude! God shall punish you for your sins!
>Gura saying her PC will blow up because it's 68 degrees >randomly my GPU reaches 92 degrees while training Had to shut it off, we're back to safe temps but goddamn this 3090 gets hot if I do overnight training, guess I have to take breaks to avoid having the thermals within being replaced... rather save that for the last few days of my RMA
>>13966 90 degrees celsius are too much for longer time but how the fuck are able to reach these temps? I train on my 3090 alot and never got over 80
>>13967 idk I had my PC on for close to a week never went above 70, after the quick rest I gave it it's 68 degrees. I honestly don't know how it reached that temps and if it was a thermal issue turning it off and on again would have made the idle temps above 50 but it's below that. Just one weird event that happened, luckily I still have RMA but I rather have the thermals replaced by the end of its date
>>13966 Gura… Those are safe temps >92 Uhhhhhh You are the one that bought used cards right? You might wanna replace the thermal pads
>>13969 I honestly think it was a bug with my thermal reporting because even with a benchmark and training with and without afterburner it still didn't go above 70, maybe windows was just acting weird cause a quick shutoff rest wouldn't fix degrading thermal pads
>>13970 Yea, if your benchmarks keep good temps despite that one incident then you are fine
>set lama-cleaner to use cuda >uses iGPU retarded dev and he wants you to pay for an installer
>>13970 is it card temp or hotspot temp, because running a specific kind of math over and over like training can push areas of the card a lot higher than general use
(2.65 MB 1536x1024 catbox_nig2k7.png)

(2.50 MB 1536x1024 catbox_k294fj.png)

(2.26 MB 1536x1024 catbox_oaypwc.png)

(2.68 MB 1536x1024 catbox_ms3eyq.png)

winged loli
>>13975 what is the hisha lora? i couldn't find it.
>>13974 I know that training usually puts more of a load on VRAM/memory temps compared to overall GPU temps, but looking at my hotspot temps they're usually under 80-85 when training, it was just a weird one time event with this 3090 but I made sure to change fan curve speeds and adjusted my undervolting to make sure I stay below 78 degrees at all times. Yet with training I only stay within the high 60s, highest I've seen before the 90 degree event was maybe 75? Otherwise yeah it'll be ok I have until next year to send this card to MSI to have them replace all the thermal pads and paste.
https://civitai.com/models/31132/based65-final-mix someone uploaded based65 to civitai, kek
(608.94 KB 576x960 493879238728097508974.png)

>Pro-China shitposting in the Image section of civitai I swear to god if we go to war with these assholes I'll enlist just so I can put them down myself.
>>13978 Well guess I have to reveal myself now
>>13976 I can't remember where I found it but here it is https://files.catbox.moe/0s4g86.safetensors
(1.90 MB 1024x1536 catbox_31h2ww.png)

(2.00 MB 1024x1536 catbox_mms1a4.png)

(2.09 MB 1024x1536 catbox_kjk6iu.png)

>>13979 Based.
>>13981 nice, thank you
>>13978 >Not mine >If guy asks me to remove it I'll take it down So literally why fucking upload or claim it. These people make no fucking sense. >>13980 I feel you
>>13982 Oh wait were u the same anon on 4chan with the ass grab prompts just now?
>>13984 I was planning to upload all the models anyways but I was busy with seeing how good Based66 is with LORAs, I still have to test out p3 epoch 8 of hll4 and see if it handles AOM3 better within the mixing recipe
>>13985 Yes, you new to /hdg/? I have basically been around since NAI leaked.
>>13987 I feel like we got at least one or two new people in the last couple day. Although I did an archive check for base64 or safe haven and it hasn't been advertised much so I guess it was through found through gayshit?
>>13989 It was uploaded on huggingface by someone else long ago but it was only used by /hdg/ anons, it's how the tripfag was able to make merges with the based models and the forbidden model.
>>13990 I meant the original base64 key to get in here but that makes sense. Also kind of glad the cock mixer tripfag isn't here after he went on a schizo tirade earlier. It's a shame since aside from fucking with THAT model, he was spelling out the block merge shit in understandable english.
(1.89 MB 1024x1536 catbox_1w33on.png)

>>13989 The more the merrier but I guess I'm just so used to people knowing my characters and my prompt "pasta".
>>13951 I had a few in my dataset and it's kinda annoying when they show up in generations, it's a waste of space
>>13991 Yeah I unironically kept a close eye on how they explained their reason for picking certain MBW presets and kept it in mind for based66, I just wish the guy wasn't such a schizo
>>13987 Yeah im quite new to the 8Chan/4Chan space. good to see u here. your gens are really nice you damn wizard!
>>13994 same, and it seems quite a few others were interested as well I did decide to read the AOM3 page after he or someone else made a pass about it in an archive post and it looks like the tripfag did take a read on what WM777 had tested with certain blocks and the anon just started schizo testing based off that. Maybe I should just compile all his posts into an obsidian document. Now if only that can help solve my issue with why I can't make a working inpaint model out of a model already merged through the standard checkpoint merge's Add Difference.
>>13995 Welcome and thanks I got too many hours put into this shit but if you need help with prompts, I'll gladly help.
(8.65 MB 3120x4096 catbox_q9t4vy.png)

>>13997 Cool I'll hop around from time to time since im not really good with nsfw prompts :D thanks for the offer
NMS just had its first actual update of 2023 so time to take a break from proompting
(1.97 MB 2048x2048 catbox_l53ld8.png)

>>13998 Anyway here's cat girl
>>14001 LMFAO HAHAHHAHA
>>14001 >pic fucking kek, but yeah I guess if they're smart enough to navigate and find this place then they pass the litmus test in my book.
(8.23 MB 2992x4096 catbox_fjjmae.png)

(7.22 MB 2736x4096 catbox_ruyq4w.png)

Heres some more..
(1.66 MB 1024x1536 catbox_9tdggw.png)

>>14004 Here is your welcome gift
>get a NaNs VAE error >I have VAE turned off holy shit, I guess latest pull Gradio really is fucked
>>14006 That came from your model. If its based on NAI it usually will break at highres fix. try the KL-F8 Anime2 to avoid that issue
(1.96 MB 1024x1536 catbox_dbmezl.png)

Asking for wariza but I keep getting good ass grabs, not gonna complain about it but I think I have used up all my good rng for today.
>>14009 10/10
>>14007 >>14008 Yea I have it already, its just that I suddenly started getting these en mass and this was the first time I got it with the vae turned off Also, is kl-f8 anime and perfectcolors the same? They look almost the same.
(1.71 MB 1536x1024 00157-3767809117-(vivid_1.png)

(1.76 MB 1536x1024 00156-3391274835-(vivid_1.png)

She's so fucking flat. God damn that's hot.
>>14009 You are collecting quite a pile of (You)s on /h/
>>14012 At least your project turned out okay, not sure how I feel about mine but her hair color is so inconsistent anyway.
>>14011 >Also, is kl-f8 anime and perfectcolors the same? They look almost the same. NTA but no, perfectolors will give you NaN errors as it's based on NAI's VAE
>>14014 >/fit/ Lorelei Mixed feelings.
(3.25 MB 1280x1920 catbox_u37a2v.png)

(3.18 MB 1280x1920 catbox_ac949h.png)

(3.07 MB 1280x1920 catbox_xe8bxn.png)

(3.50 MB 1280x1920 catbox_hhlv90.png)

Out of curiosity I tried to do a low dim Nahida using OG Nahida anon's dataset with updated tags and my current low dim training settings. Unsurprisingly not as good as yetanothernahida but it does try to do something with the eyes.
>tried doing 1024x1024 on 3090 batch 4 >gpu blows up and mustard gas starts flowing out of my pc 48gb vram gamimg gpu when
>>14019 just nvlink
>>14020 I thought NVLINK didn't combine VRAM for LORA training
what's the fucking tag for a "from the front" angle, even pov doesn't work
>>14022 "facing viewer" might help if it's a person you want facing towards you
>>14023 nope, same thing as "looking at viewer" pretty much i want an angle like this, literally straight on, nut turned slightly or slightly from the side, literally straight on there must be a tag for this, i refuse having to use img2img or controlnet for something this simple come on..
>>14001 I removed the post because pictures of people are strictly not allowed on this site but I'm not issuing any ban/warn/whatever because I can understand it's literally just a joke/meme. Just posting it in case anyone's wondering 'why did this get deleted?'
>>14024 hmm, looking at danbooru I don't actually see a tag specifically for that. the only tags I found that sort of correspond are "facing viewer" and "looking at viewer", but both of those tags have a lot of images where they aren't facing directly at the viewer. might have to neg some of the other nearby views, like side view or looking to side
>>14024 Potrait, straight on, cowboy shot?
Does loading models depend on the hard drive you are using? At this rate it's the only reason I'll use a gen 5 NVME
>making gens for the Basedmix demonstrations >65 protomix gets LORAs details better than finalmix yeah it must have been one of the finetuned models I mergeblock weighted holy kek
>>14029 cpu + drive + gpu bandwidth i'd imagine drive speed probably doesn't matter above sata ssd speeds tbh
>>14027 >Spend 300-1000 on gpu >Can't afford a 50$ drive
>>14029 PCIe slots speeds will trump sata drive speeds unless you are using a caching drive to speed the latter up.
>>14030 It doesn't seem to like merging multiple finetunes does it.
>>14027 Crucial P3 plus 4TB drive is on sale for 200 bucks
(1.19 MB 1120x1120 04972-2518602214.png)

>>14029 yes. This is my loading speed for a FP16 pruned model on WSL2 using a Samsung 980 Pro
>>14035 Yeah guess if I decide to go with the different versions for Based66 or go straight to based67 for HLL4 P3 epoch 8 it will be the only fintuned model I keep within the recipe
>>14024 try "focused"
>>14041 I was wondering what this is but it's an object remover, this is cool I wonder if it works well with Anime stuff
>>14039 that's a shame because I would've loved to make a personal basedmix with hll and my finetune
>>14043 Honestly I'm not that autistic with block weights yet, final-mix for 65 was the first time I used it and I was trying to understand tripfag's posts and the rentry explaining it. I bet you that a more expert model block merger can work with multiple finetuned models and still ensure LORA compatibility
>>14044 might need to chat up the people in /vtai/ because they make lots of mixes.
>theresa is my litmus test to release the next based mix yeah based66 mix needs more work with LORA compatibility before I release it, at least I have a good recipe route to get the final result for that sharp line anime look I got before. >>14045 Yeah I think I'll head over there once I start rebaking Based66, I want to see how the model mixers there are doing stuff.
>>14046 Actually doesn’t one of the AOM3 versions use a couple finetune models? Might be a good control variable for testing if it is the case.
>>14047 I was using AOM3A1B for based66, I'll have to look over again to see which one is the finetuned model
>>14048 So according to what I'm reading, AOM3 uses Counterfeit as one of their models,and both the huggingface and the Civitai page say its a trained checkpoint, with the 2.0 paging specifically mentioning its a trained dreambooth thats been MBW and merged with LoRAs. https://huggingface.co/gsdf/Counterfeit-V2.0 https://huggingface.co/gsdf/Counterfeit-V2.5 https://civitai.com/models/4468/counterfeit-v25
>>14049 Ah... So technically for Based65 and 66 I was using 3 finetuned models, holy kek no wonder why LORA compatibility was weird for those two compared to protomix and based64
>>14050 I tried playing around with Counterfeit for mixing and I tended to have some issues with it with how my gens were turning out. I also seem to recall people didn't like how AOM3 was reacting to LoRAs so I suspect that may be the culprit. The AOM3 core model is essentially AOM2 Hard but the Add Difference processes to add Gape was upgraded from the simple checkpoint merge to a BWM. So maybe try using AOM2h or AOM3 and see if you get better results.
>>13975 > <lora:hisha:0.45:face> What is this syntax? Is there an extension that let you divide and name regions then apply lora to just a region?
>>14051 Yeah that sounds like a good structure, I'll have to see if it works well with the recipe structure I had for Based66. Otherwise given finetuned models I believe going above 2 finetuned models within a mix will start making LORA compatibility wonky.
>>14053 Hope to have been some help. I'd say that if you immediately see improvements using the non-counterfeit mixed AOMs, that maybe we should avoid counterfeit and possibly dreambooth models.
>>14054 Yeah, if anything 2 is my limit but now I have to look up some of the models in my recipes to make sure they're not finetuned models because the only one I'll be sure to use is HLL3 or 4
(4.18 MB 3968x3941 tmpe3okn1wn.jpg)

>>13889 It's hilarious how this lora exposes more and more forehead at higher weights
>>14055 I would probably look at it more like, if you can’t identify how a model was trained and if it wasn’t trained on NAI or SD1.x, it’s probably a no go. AOM uses realism models that are very likely finetunes themselves (F222, Cafe-Instagram, Basil). Most new models tend to also merge LoRAs and those get harder to identify if they were trained on NAI or they did something crazy like use AnyV3 or AOM as the base ckpt. If there is no paper trail to track the ingredients, you have a judgement call to make or just experiment for the answer.
>>14057 True, I guess I'll have to put info on my upload pages to tell people to be wary of mixing any of the based mixes with other models once I mass upload them on civitai and huggingface. I'm just trying to get "good" preview images for each model before I upload everything.
(1.79 MB 1024x1536 catbox_kz5tst.png)

>>14056 yamamoto-san knows what the fans want 0.6 takehito harada fixes the hands up pretty well
>>14058 I’m awaiting your future success. Also have you reported that thieving nigger that uploaded your shit?
>>14060 Just told him he could keep it up until I upload everything kek
>>14061 but looking at "everything" holy shit Based64 mix 1 literally didn't do what I originally wanted, it's just AOM2's hard look but more intense and that ruins the purpose of me trying to get rid of the basil look
(1.68 MB 1024x1536 catbox_hw4f1c.png)

(1.47 MB 1024x1536 catbox_1pkzkd.png)

(1.69 MB 1024x1536 catbox_h4wni5.png)

>>14056 >>14059 The forehead is where the love is stored the bigger it is the more lovely is the girl
>>14063 you must love Ayaka
>>14063 is that lora released?
(1.58 MB 1024x1536 catbox_up1zyo.png)

>>14064 help an anon out as there a too many ayaka chars/artist >>14065 https://anonfiles.com/Ccv5qejaze/haguhagu_rinjuu_circus_safetensors
(1.64 MB 1024x1536 00041-3766315575.png)

(2.00 MB 1024x1536 00010-3026873708.png)

(1.71 MB 1024x1536 kotoyama-788it-novae.png)

(2.15 MB 1024x1536 00053-2044635513.png)

https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg kotoyama (dagashi kashi mangaka) added
(1.83 MB 1024x1536 catbox_3a5929.png)

(1.85 MB 1024x1536 catbox_i2i1qx.png)

(2.31 MB 1536x1024 catbox_c2cro5.png)

(2.20 MB 1536x1024 catbox_6uj8ki.png)

>>14052 Hello ye this is made by using this extension: https://github.com/hako-mikan/sd-webui-lora-block-weight With this, you can apply lora weight to only specific layers. Some presets to play around with: face:1,0,0,0,0,0,0,0,0.8,1,1,0.2,0,0,0,0,0 Nface:1,1,1,1,1,1,0.2,1,0.2,0,0,0.8,1,1,1,1,1 ware:1,1,1,1,1,0,0.2,0,0.8,1,1,0.2,0,0,0,0,0 pose:1,0,0,0,0,0,0.2,1,1,1,0,0,0,0,0,0,0 paint:1,0,0,0,0,0,0,0,0,0,0,0.8,1,1,1,1,1 char:1,1,1,1,1,0,0,0,1,1,1,1,1,1,1,0,0 back:1,1,1,1,1,1,0.2,1,0.2,0,0,0.8,1,1,1,0,0
>>14066 >>14067 what a nice day, thank you both
>>14062 Yea it sounded like at that point you needed to remix AOM from scratch without basil or chillout (since it uses basil) and probably go back to AOM1 which is F222+Cafe Instagram for the realism or find a better base realism model that isn’t the overdone bullshit you see on civitai like the mega porn model or some shit.
>>13647 >>13644 Which sakura lora is this? I don't see it on gayshit, is it the one on civitai?
Does anyone know what the small hexagon artifacts on image generations mean? Is my GPU pickled?
So Civitai has a retarded way to upload models >wanted to use one page to upload all Basedmixes and future versions >nope have to make separate pages and when I try uploading Based64 V1 and V3 Civitai whines about them being the same file format and size pain
(1.76 MB 1024x1536 00003-2296776832.png)

(2.04 MB 1024x1536 00008-3026873708.png)

(2.24 MB 1024x1536 00052-2044635516.png)

(3.72 MB 1536x2304 00038-3671405465.png)

>>14073 upload it to huggingface and call it a day. infinitely less pozzed than civitai anyways
>>14071 do you need a sakura lora? i haven't found an anime model yet that doesn't respond well to kinomoto sakura, cardcaptor sakura.
>>14075 Probably not, but there was one in the metadata of those images.
Reposting these from ATF, (yes I know honey pot site blah blah blah FBI agent watching me jerk off) but they have an AI community and they're the only other site I know of making uoh-LoRAs maybe itll give you guys some ideas if you wanna bake better versions or something idk. https://gofile.io/d/615nd2
>>14077 no one here has an issue with ATF. you're not on 4chan /hdg/
(1.50 MB 1024x1536 catbox_pat1qv.png)

>>14077 they uploaded a haguhagu lora a few days ago coincidences are funny
>>14077 Huh, I haven't checked that. How are the quality of those compared to others? Like they even have an Asanagi, does it beat the others that people currently use?
>>14078 i have an issue with a number of their users for being really stupid and annoying
(1.90 MB 1024x1536 catbox_7yt0of.png)

(2.12 MB 1024x1536 catbox_3j271c.png)

(2.00 MB 1024x1536 catbox_asmwvk.png)

(1.99 MB 1024x1536 catbox_yr7le6.png)

uooooooohverdosing again
>>14077 >yassy *unzip*
(1.10 MB 768x1152 g3o0QcY.png)

Has anyone managed to snatch Leotard Lora before it got nuked from Civitai? Link was https://civitai.com/models/27762
(2.13 MB 1568x1024 Haga Yui (1).png)

(2.19 MB 1568x1024 Natis1.png)

(2.29 MB 1568x1024 Quzilax (3).png)

(2.58 MB 1568x1024 Haguhagu (2).png)

>>14081 yea they're kinda stupid they hardly know what they're doing atm when it comes to LoRAs I think they've only gone as far as doing decent merges (berry mix and bstab originated from there) but as far as LoRA training they're barely getting into it where as 4chin and here already have done multiple LoRAs to a decent success. >>14080 on about the same level of civitai quality loRAs, not sure I haven't messed with them myself, but im attaching some of their preview pics just so you can see.
>>14085 Why did you post that second one you fucking nigger
>>14085 The faces on that 2nd one are nightmare inducing
(1.19 MB 1024x1024 Yam Nai.png)

(636.04 KB 672x864 salad nai.png)

(784.24 KB 672x864 Takaji Nai.png)

(924.94 KB 672x864 Maka Fushigi nai.png)

>>14086 because im fucking retarded and I didn't look at the image before posting it, yeah here's some more LoRAs (with more previews attached), im not posting anymore after this cause im too lazy and they might suck. Benantoka https://store6.gofile.io/download/c8e203ee-6144-4722-aa60-b88946a10c0a/BeNantoka Nai.safetensors Artist: Maka Fushigi. https://store1.gofile.io/download/cfe7acac-bcfb-42dc-8f3c-b08e36c82e40/Maka Fushigi nai.safetensors Miyasaka Takaji https://store6.gofile.io/download/1737f933-4e81-4eed-8857-3a485e14dba4/Takaji Nai.safetensors Salad. https://store6.gofile.io/download/df4224e2-8ce8-48c3-a6e1-04eebe6f6e6d/salad nai.safetensors Yam https://file80.gofile.io/download/2360431b-9df9-49ad-828d-5b5906cbbe73/Yam Nai.safetensors
>>14084 yeah I snatched a copy before it got nuked I think this is it, not sure it was named like this in my recent dls so its probably it. https://anonfiles.com/Wbffudjaz5/leota_safetensors
>>14088 What the hell, looks like they just got nuked
>>14088 I'm not impressed at all
>>13958 seraziel lora v5 cooking
do you add BREAK even on single block prompts? is it placebo?
>>14092 i admire your dedication anon crossing my fingers for the knots
(1.91 MB 1024x1536 catbox_otc8tf.png)

>>14085 >>14088 these are pretty cursed
>>14090 >>14095 Yeah they are I think cause I can't dl any either mb on posting the dead links then, anyways I just wanted to give you guys some inspiration if you wanted to tackle any of the artists cause clearly these are just poorly made. I'm not gonna throw shade cause I don't train LoRAs myself but im pretty sure they can be done better
>>14089 Thank you, anon!
(2.01 MB 1024x1536 catbox_mt1zkp.png)

>>14085 while I like comparing my shit loras to others these are beyond cursed
>>14100 got an example? Yassy is weird one where I love the eyes but get weirded out but certain things he tries
>>14101 those WERE the examples, it seems to be a bit undercooked because the style resemblance isn't super strong but the LoRA still looks nice https://files.catbox.moe/bmgfpw.png
(1.57 MB 1024x1536 catbox_d96zf0.png)

(1.60 MB 1024x1536 catbox_2bgozk.png)

>>14102 holy shit Undercooked is an understatement https://gelbooru.com/index.php?page=post&s=list&tags=yassy+&pid=0 These are pics from him... dafuq did they do?
>>14103 I dunno, they clearly trained it enough to pick up a style because it looks totally different from based64-v3 default, but yeah they must have had a wacky dataset to get results like that
>>14104 actually looking at the metadata I can see the issue, yassy has maybe 100 usable color illustrations for a dataset, this was trained with batch 12, 229 batches per epoch, 11 epochs with 1e-4 lr and no txt encoder lr set... 512x512 too so either this was a very early LoRA or they're still training LoRAs with january recommended settings
>>14085 >>14088 Thanks for saving me time I would have wasted downloading and trying those.
>>14041 I'm a bit lost on how this is "amazing", I thought segmentation had gotten really good in the past few years already.
>>14077 Already taken down?
>>14085 kek 2nd one gave me a good laugh
swin or vit for tagging?
>>14110 I always use convnext but I've never done a thorough comparison, nor have I ever seen a good comparison by anyone else.
>>14110 the creator of the model recommends swin
>>14110 I always use swinv2-v2 not git, but dunno if it's the best
I guess occasionally we get to have some 4chan level action too lol
>>14085 i don't mean regarding loras, i mean as people. a couple of them posted in the /b/ threads and they were some of the most contemptible retards i've met in my life. impulsive, low iq, only visible talent was for retreating into emotionality and equivocation to escape blame for their own actions. genuine niggerbrains.
>>14116 you dont sound like a reasonable person yourself
>>14117 there's nothing unreasonable about hating people like that.
>>14117 He's got a point, certainly not the kind of people I want here
(2.57 MB 1280x1920 catbox_z09com.png)

(2.35 MB 1280x1920 catbox_lup721.png)

(2.59 MB 1280x1920 catbox_jdkx5h.png)

(2.45 MB 1280x1920 catbox_cqgkhd.png)

>>14116 yeah that too tbh, honestly I don't ever post there and just lurk and steal prompts from time to time (but you have to dig through a ton of bad gens to find good ones), I don't join the circlejerk because I prefer the anonymity imageboards provide.
>>14094 red rocket seems to works natively now... sometimes It can do regular penises too but it can be finicky, especially on certain seeds or on furry bodies. I'll put it back into the oven for a bit since it's "only" 8k steps compared to 14k for the v4. It feels undercooked but seeing the epoch comparisons it probably landed on a restart. I'll hope it'll fix those problems a bit
Still wondering if getting hexagon artifacts on outputs is a sign of a GPU defectiveness, by "hexagon artifacts" I mean when you zoom into an AI image and you notice hexagon shaped artifacts that look a bit fuzzy/noisy, it's always near the bottom of the image
(11.08 KB 344x113 1ckpb3.JPG)

>>14123 just in case this might seem like a schizo disillusion
>>14123 >>14124 I've never heard of this before.
>>14124 >>14125 It doesn't seem to happen on every generation, I notice it only on a few generations of mine that have more focus on darker shades
>>14126 yeah majority are clean, idk what is the issue
>>14127 Maybe the xformers I built with the latest dev build for torch? I have to install the xformers I made with 2.0.0 to make sure
>>14124 looks like it's struggling to reconcile (detailed skin:1.3) with either bokeh or a model that loves blur vignette
>>14129 Probably I have a few ways to test out the specific seed to see if it’s the gpu itself but seeing it’s just 1/10 of the outputs I looked at maybe that can be the case, VAE can also be a culprit
>>14122 >added min_snr_gamma because why not >launch the bake >play warblade (because I can't game and train lmao) >check back 40 minutes later >it's actually 2x slower >have to restart to have it take a regular amount of time From 4s/it to 7.5s/it (batch 4 dim 128 size 768 grad checkpointing) probably isn't worth the possible added quality
(9.87 MB 2736x4096 catbox_wehycg.png)

Posting some of my old gens
>>14130 Yeah it seems to be VAEs that try to pretend to be something new but end up being NAI.vae that cause most of those grainy artifacts
>>14124 I get these too not very often but it happens, probably because I use >ayyymd (yeah I bet on the wrong horse im a retard) but it does happen on my end sometimes,
(1.62 MB 1024x1536 catbox_k4kqqe.png)

>>14105 I am currently working on my version but it seems a lot harder than I thought gotta rebake it till my autism is satisfied
>>14100 Yeah that's pretty cute, I guess I'm giving it a try.
>>14134 >using NVIDIA tho I wonder, I might do a clean reinstall of my GPU just in case but the VAE I was using was NAI based and the Waifudiffusion 1.4 VAE barely cases such strong noisy artifacts like that
>>14135 that looks a lot better already if you really want to get into the woods, he's released a number of full-color doujinshi
>>14016 I am spiritually and morally obligated by the greater will to make as many /fit/ anime girls as I can.
>>14131 I have same gen speeds with it on, quality I don't know about but it's not hurting anything for me far as I can tell.
>>14140 Yeah it's probably not that. It just seems like I have cursed luck when it comes to this, sometimes it takes a little less than 10 minutes to reach around 4.5it/s cruise speed, but sometimes the plateau is 6-8it/s. I thought resuming the lora made it like that but it just worked albeit without snr_min_gamma, and I'm not confident enough in my luck to just restart it and pray it goes well. Does it actually make it better? Does it need further tuning?
>>14141 I meant s/it not it/s otherwise I'd have the fastest 2060 in the world
>>14138 I already scouted the doujins and cg sets which is why I will never understand what a man has to be tormented with to create over 400 variation of the same picture...
I can finally do breast envy stuff, based regional prompter.
(1.82 MB 1024x1536 catbox_1fziqb.png)

(1.39 MB 1024x1536 catbox_dkqwmn.png)

(1.61 MB 1024x1536 catbox_hy6oda.png)

(1.66 MB 1024x1536 catbox_9irbt6.png)

Making progress
>>14145 Might look more accurate if you specify flat shading in the prompt and 3d in the negative? These certainly don't look bad by any means though. I really do feel like AI sometimes can produce a "what if this artist refined their style more" for lack of a better term
(1.74 MB 1024x1536 catbox_1zkaml.png)

(1.69 MB 1024x1536 catbox_fi5jmg.png)

(1.73 MB 1024x1536 catbox_12ox91.png)

>>14146 As I need sleep I will give you the current state maybe you can prompt better than me https://anonfiles.com/T3z206j2z0/yassy_safetensors
>>14124 It might be artifacts from the diffusion program but on a possibly related note, there are some images that use a kind of dotted pattern (usually in steam clouds), and I wonder if that could have had some effect in the training data. Examples: https://www.pixiv.net/en/artworks/103286229 https://www.pixiv.net/en/artworks/92365636 I'm not an artist but is this a common technique or some preset brush in programs like Photoshop? I've seen it multiple times. Additionally there's probably all kinds of grainy/low-quality/low-lighting photos in Stable Diffusion as well that could mess things up.
>>14148 Yeah that actually makes sense, but if there were actual issues with the GPU how would that affect outputs? Because from what I understand if there was something wrong with the GPU the bad effects would be drastic and not something small like this that only happens 1/10 times
>>14145 >>14147 looking pretty good, time to make yassy futa doujins with AI kinda makes me want to train frogsnake but I'm pretty sure any attempt to do so would just feel off unless I specifically trained their character depictions
>>13937 I've become so desensitized by prooompting that the first thing I notice in these pictures is that absolutely beautiful and coherent background. God damn. And here I am stuck with muddled grass. Toph is also nice I guess. I apologize in advance for this nonflat hag, but I'm just showing what I mean with the background. Most loras I used seem to mess with background fidelity.
>>13806 is that Suwako a LORA?
Is there an up to date repository or at least list of what Civitai has shoah'd?
>>14153 Not really. The only one I’m aware of is the motherload rentry on /g/, in the sense that it civitai nuked something, you should probably check there since the rentry owner has made it his mission to reup stuff there.
>>14148 nah I still see it with the same seed on my 2nd 3090 too >inb4 I have to switch with my 3080ti to see if it does the same thing not worth it I'm just going to believe that it's the model and whatever you said outputting generations like that sometimes.
(1.63 MB 1536x1536 catbox_th4zxl.png)

>>14154 Thank you for the spoonfeed anon. our entire operation would be nothing without reentry in general
>>14155 even shows up on the same seed with torch 2.0.0 and the xformers I built with 2.0.0, maybe I should go schizo and pop in my 3080ti to see if it does the same thing but honestly with these tests it just seems to conclude with >Additionally there's probably all kinds of grainy/low-quality/low-lighting photos in Stable Diffusion as well that could mess things up.
>haven't done any finetune prep in two weeks >still have to add more images and clean tags Why did I choose the most autistic path...
>>14157 I went full schizo and put my 3080ti in place of my 3090 and that grainy noisy artifacts on the lower part of the same seed output appears too, also holy shit I forgot that the 3080ti actually used more watts than the 3090
(1.85 MB 1024x1536 catbox_jm3h8w.png)

(2.10 MB 1024x1536 catbox_4ztppx.png)

(1.90 MB 1024x1536 catbox_1xo8ah.png)

Stupid cute butt and belly
>>14160 i've been doing a lot of butt shots lately they are very nice when you find a style combo that gives puffy vulva from the back
>>14160 fucking hell anon your gens evolved from "good" to "amazing" as time goes on
Does anyone have info on Kawaii 2D/2.5D? Wanted to if there was a recipe mix for it but the Civitai page was nuked.
>>14163 Maybe someone on 4chan Hdg archived it, go check I want to take a look at it myself
>>14163 theres an anon who scrapes metadata from civitai (check the last thread) on here maybe he has info on it
>>14163 >kawaii Do you mean the animelike 2D/2.5D models? There are two models on huggingface that look like reuploads, as well as one in https://rentry.org/sdg-motherload#animelike-2d-v2 , but I don't know if they have any training data in them. I only got the two lora some anon extracted from the models, which are up on gayshit.
Back with more image enhancement for datasets. This time, denoising and I wouldn't really recommend it unless you really hate noise and don't mind losing fine detail. From weakest to strongest: Original -> ISO_denoise_v2 -> ISO_denoise_v1 -> NoiseToner-Poisson-Detailed
>>14168 there's a lot of great models there, gotta make sure which one has a fintuned model within their mix if I plan on using them for the future Based mixes
>>14168 Ah, nice. My eyes immediately jumped to the animelike model and since they're separated in the table I didn't notice the other two
>>14169 I'm doing essentially the same thing, got to find out what everything is made out of or else you get this >Differencing SD/NAI out of Counterfeit
>>14169 Anon are u the based model's author? if so i have a question. Do u use any of OrangeMixs in your model? aswell as block merging?
>Try to inpaint away some messed up folded hands >End up with a phone instead Eh I'll take it.
>>14172 Not him but it was discussed earlier in the thread about AOM usage and block merging in the based mixes.
>>14173 Happy accidents lol
>>14173 mistakes into miracles
>>14172 AOM2 hard was the base for 64 and 65, AOM3 is what I started using for Based66 but I might revert to AOM2 for based66 after finding out AOM3 uses counterfit 2.0 which is a finetuned model and too many finetuned models within a mix conflict with LORA compatibility
(1.32 MB 864x1304 06832-949906100.png)

>want to get 1300w >there’s a 1200w psu on sale that has two 12HVP ports Not sure if going for that will be a bad decision
>>14179 I use a 1200w with my 4090, what are you trying to set up?
>>14180 Just wanted to future proof if I put in a 4000 series with a 5000 series card
>>14177 I see. do u have an archive where ican download all the based models? im a block merger myself and I could use some of your models as a block merge.
>>14182 Just remember to avoid based65 final mix the last two mixes I put in there compared to protomix have way more than 2 fine tuned models so it conflicts with LORA compatibility https://huggingface.co/AnonymousM/Based-mixes/tree/main Or you can try to be more autistic than my first attempt with MBW presets and trying making more than 3 fine tuned models in a mix work better with LORAs, gl
>>14181 Two 4090s take about a little over 1200w, maybe 1350 will be more safer… ugh I hate all the new power connectors added to this generation
>>14183 I'm sure you know more abt block weight merging more than I do but would u like some resources? https://www.figma.com/file/1JYEljsTwm6qRwR665yI7w/Merging-lab%E3%80%8CHosioka-Fork%E3%80%8D?node-id=1-69&t=Npij3JkbvnRyqWQW-0
>>14185 Yeah that’s useful, thanks anon
>>14186 If u ever want to add more info to this pls mention the word. Baka Man and ill show up. LOL
>>14185 NTA but nice
>>14185 >see lots of hands neat
People bring up vtubers so much that some have started to live rent free inside my head Like that one clip of shondo screaming ITS FUCKING OVER AHHHHHHHHHHHHHHHHHHHHH like an unhinged banshee whenever anyone says its over
>>14185 how do I save the pic from the site?
>>14191 clone this to your own account there should be a button somewhere. then u can edit everything and should be able to save files
>>14190 I have a similar feeling except I get nam flashbacks >See Kronii Uno livestream on my tab with Mori and Tempus >Imagining the Kronie schizo sperging out on his twitter and other anons crossposting global and /infinity/ about males and getting banned Yea I'm glad I only visit /vtai/ these days.
>seeing the document updated in real time this is fucking cool
>>14194 Through dick, unity
>>14194 I'll be adding more info in the upcoming few hours. but yeah ill refrain from doing it in realtime for now and push an update later.
Pretty new to all AI shit, but I recently finished my first LORA. Works great, but part of me thinks it was more dumb luck than anything else. Wanna make a few more about some esoteric concepts, but I feel like I only have a very rough understanding of how to properly tag the images. Can anyone direct me to a good resource in regards to this stuff? Or just explain it. Whatever.
>>14196 its totally fine, do what you are doing even if its live. On a side note, you wouldn't happen to be able to source information for finetune training?
>>14198 I do have information on finetuning too but finetuning a fullmodel is quite different from LoRAs or LoCons if you're used to them. check the EveryDream2Trainer from github they posted and explained what most settings do.
>>14199 I've already finetuned using Kohya before. I have the EveryDream2 page bookmarked but haven't used them yet. But I know for sure my finetune is not optimally trained.
>>14200 I've only used StableTuner for fullmodel finetuning. and its quite lacking in features so im not sure how different Kohya or EveryDream is compared to Stable Tuner. the Stable tuner project is feels dead.
>>14201 It seems EveryDream is the most active one while Kohya is mainly focusing on the LoRA and co. tech and just at least maintains the finetune trainer.
>>14202 Yeah go with EveryDream anon. its better
>>14203 I'll probably jump on Everydream after I finish these last test trainings since I don't want to fuck up my consistency but I have been considering it for the future. And if it works out, its not like my work goes to waste.
Should I dig up that tripfags posts about his MBW experimenting?
>>14205 if its useful info I can add to the Figma sure ill take a look
>>14206 I'll copy paste his posts into a rentry and post it a little later tonight then, and just scrub mentions of the forbidden model
>>14207 Cool. just mention the word Baka Man and ill appear. Probably. but i need some sleep too so gotta go now. thanks for the info
>>14208 sounds good, I'll just reply to this post again
>>14197 You probably need to check the booru tags to get the concept that you wanted so you can easily pruning tags into your concept.
(26.45 KB 488x398 export-in-figma.jpeg)

>play modern game with max settings >uses more power than training I wonder what directx13 is going to do to older GPUs
Does there happen to be any interracial/blacked models or loras?
>>14213 go back
>>14151 don't apologize for nonflat gens! big and small tiddies are beautiful and should be gazed upon! >>14152 well... it is my first try with a locon https://anonfiles.com/9ebe72j4z9/average_hanzoLoConV2_safetensors not that good but workable
(1.37 MB 480x360 desktop.webm)

>>14214 based and blackedpilled
(1.41 MB 1024x1536 catbox_fh7p57.png)

(1.87 MB 1024x1536 catbox_twv6xp.png)

(1.86 MB 1024x1536 catbox_6s2sfu.png)

(1.67 MB 1024x1536 catbox_hychlh.png)

>>14216 some example of the locon my problem is that the hands are even more fucked up than usual
AI art should be a tag on pixiv instead of a global setting because I want to disable AI art for certain tags like Nahida but want to see it for other tags
>>14217 >this one spergs out on penises BROS I think I'm gonna end up trying to block-merge v5 and v5.2. Anyone have info on that for loras?
>>14221 Okay I found a way to test block weights on LoRA, but is there any way to merge ?
>>14222 I'm on the verge of just doing a script that generates all possible images for a set of weights to do an exhaustive search
>>14222 https://github.com/hako-mikan/sd-webui-supermerger/ never used it myself, but I've heard great things about it and it looks like it should work for LoRAs
I still wonder if doing batch 3 is worth it, I can do it with “1 megapixel” LORAs but maybe 2 is more worth it >saying this when our best LORA maker can only do 1 batch I’ve really fallen behind with my LORA skills desu
>>14225 >saying this when our best LORA maker can only do 1 batch is there a ranking list? And if yes is the shittiest lora the last or first place?
>>14226 Just a personal ranking from final output and the anons that have uploaded all their LORAs to one place. No one here makes shitty LORAs, last place would be people taking YouTube advice
Vtuber shit but all of Kawaii pro gen 3 (the remaining members) will be terminated due to disputes that couldn't reach an agreement with management, their alts if you care https://twitter.com/EliaStellaria Aletta https://twitter.com/AiriViridis Peony https://twitter.com/GezuDayo Sava https://twitter.com/kirispicaa
>>14228 ah forgot, the last link is for MIryu and Oceane is right here https://www.youtube.com/@UtadaZoey
(2.10 MB 1024x1536 catbox_49fk48.png)

>>14228 I wonder how many vtuber are existing right now lost count ages ago
>>14230 millions at this point, even if AI vtubers entered the market it's already over saturated ever since people wanted to get a piece of HoloEN's fame. Kawaiipro was good for cute girls GFE for me but ever since gen 2 they got more of the "self-righteous yassssss girl" type of members so honestly after 2021 a lot of over-saturation with vtubers like this existed, getting GFE even if it's kayfabe is impossible in the EN market until there's a dedicated AI vtuber group that leans into the >"my AI girlfriend" market
(1.33 MB 1024x1536 catbox_mkw8ko.png)

>>14231 time to make your own vtuber and use ai technology to make her as real as possible
>>14232 it's possible I just need to know about the open-source tech for Live2D and get an open-source for AI voices
>>14233 and for language models I believe it's possible to run Lama on a 24gb card but I need to dig more into open source language models to run for an AI vtuber to avoid censorship, I want her to be as based and hot as AI can be
>>14234 but it gets technical when you put in training an AI to play video games too, Neuro-sama is fascinating knowing how she was just an OSU gamer AI turned into a full fledged Vtuber
(1.62 MB 1024x1536 catbox_6n775u.png)

>>14235 we have the technology! we just have to use it and create the perfect waifu the biggest hurdle would be open source language+voice Share your progress anon... if you are insane enough to do it
>>14236 I might once I can focus on looking up everything, guy that made Neuro-sama should make everything they do to make her work open-source desu
>>14227 oof, and here I was thinking I was doing a good job with my loras, to be fair though, I don't ever see anybody using them... so perhaps I am bad at making lora and just tricked myself into thinking I was good. oh well, such is life
>>14224 Nice. I guess I'll have to do the tests manually too since I can't just do X and 1-X to merge blocks
(1.75 MB 1024x1536 catbox_u5esva.png)

>>14238 it's not about seeking validation it's about meme potential or nice booba my lora are very rarely used but I love shitting them out nonetheless
>>14240 I have quite a kek when I see your OC
>>14238 I don't know which LORAs you post but I barely use other people LORAs too, I just archive them the only time I have used LORAs is to post outputs of mine to see if the LORA I made is satisfactory to me before posting it here and on MEGA/CIvitai. It's more about validation from my baking process and seeing if the parameters are using give me the "best" results.
(1.67 MB 1024x1536 catbox_w8m5jm.png)

>>14242 same I even remade lora when the one found was not up to my personal standards Are my lora better? Oh hell no but I like to think that I get better by cleaning training data and trying for myself
>>14225 >saying this when our best LORA maker can only do 1 batch >our lol
>>14243 Yeah whenever I see a LORA not get exact details or fail the (different look:1.4) test I just throw it out, look through my parameters and look through the training data to see if I can make the LORA better to my standards, even with a 3090 I still spend as much time as I did with my 12gb card baking stuff
>>14240 true, not that i've had time to bake lora or even gen the last few days. the update to sd-scripts is making me think about the logistics of the popups and it's becoming increasingly annoying to make a popup for. >>14242 Fair! you have a point, but hey, the more styles that get made, the more combinations that 'can be made. I'm actually working with my group on making wildcards for styles as part of the wildcard system, but we have a logistical issue that we have to figure out, namely, how we are going to be able to distribute the style lora we use for the styles. after all, we don't exactly want to step on anybody's toes by re-hosting their lora, but we do'' want to let people have this really cool set of styles that we have started concocting.
(1.77 MB 1024x1536 catbox_593qqj.png)

>>14245 now that sounds interesting what is the (different look:1.4) test ?
>>14247 If you have a character LORA and if their hair style and outfit can be changed with 1.2-1.4 and still get their main look that is a great character LORA for me
>>14248 Style LORAs are easier in that sense, the only thing I look in for style LORAs is seeing if merging it with different LORAs only requires a 0.8 strength
>>14219 I thought about training averagehanzo, did you run the source data through an upscaler or did the AI just not learn the pixelated lines?
>>14250 oh no it learned it just right which is why I said it fucked certain things like hands more up than usual I trained several tries and merged them with each other till it was... the current locon not perfect but hey if even one anon likes it it can be a win in my book
>>14221 Read something interesting on /hdg/ Since I have a few concepts I want it to learn along the style, I should probably up the TEnc LR. Should I try upping LRs from 2e-4/1e-4 to 2e-4/2e-4? Might even try 2e-4/3e-4
>>14252 Also, did anyone tinker with overlapping tags/tokens? For example, if I have "penis", "human penis" and "horse penis" in tags, how much impact is there from A. The penis tag being a prerequisite tag for both and potentially muddling the concept B. The "penis" token being part of both tags, which could maybe fuckup something on a more fundamental level (second one is mostly schizo since I can't get it to work AAAAAAAAA)
(1.45 MB 1152x1152 catbox_9wes3j.png)

(1.52 MB 1152x1152 catbox_zeajwv.png)

(1.48 MB 1152x1152 catbox_7n4vuh.png)

(1.25 MB 1152x960 catbox_d8qcul.png)

>>14251 Oh yeah I see now, it does do the linework pretty faithfully if you use a non-latent upscaler. I can see what you mean by fucked up hands, these are the best examples out of doing 4-batch gens of each character lmao. Still very good and I'll probably play with style mixing with other chunky artstyles.
>managed to get hot spot temps under 80 degrees when baking LORAs trying to find good clock speeds and voltage spots for an originally power hungry GPU is fun
(1.80 MB 1024x1536 catbox_paizci.png)

(2.24 MB 1024x1536 catbox_04sbp4.png)

(1.67 MB 1024x1536 catbox_j83hlc.png)

(2.00 MB 1024x1536 catbox_4msrp3.png)

>>14254 which is why I was hesitant to upload it but maybe it is salvageable if you mix it with other styles
>>14255 I don't have to worry about that since I'm on a laptop, everything is bound by thermals (and powers, which upon touching seems to also be thermally bound). I just have to turn on hybrid mode to have regular stuff on iGPU, and unplug any external monitor, otherwise speed suffers a lot.
>>14257 >touching the power brick I mean not a lot of sleep these past few days
>>14256 I'm of the opinion that there's no harm in sharing something flawed because SOMEONE will either find a use for it or use it as a launchpad to make a better version. I think I'm up to like 3 LoRAs I made to try and improve on other peoples' that I wasn't quite happy with.
>>14259 Yeah there’s no harm but honestly does sting when someone else manages to perfect something you did many bakes on, at least I know anons here would share what they did different but being on the spectrum gives you a lot of stupid ego
(1.78 MB 1024x1536 catbox_lenk5h.png)

>>14260 >>14262 oh no someone stole the clothes of this nice ladies who would do such a thing?
maidanon what is the funny variant of DPM++ 2M Karras you're using
>>14265 A tweaked version I got from this anon: >>13466
>>14266 >>14267 the sampler had a bug and it was a ONE VARIABLE change? that's hilarious, I will definitely run some tests with this, thanks
(1.96 MB 1024x1536 catbox_jsaukr.png)

>>14267 how much is tweaked? I am currently trying my luck with it
>>14269 From the testing I've done makes it seem like it is actually sharper.
I noticed that in some prompts I get a slightly washed out color even with vaes on
(585.14 KB 2304x2304 catbox_4qnq8c.jpg)

(566.36 KB 2304x2304 catbox_vcrv3p.jpg)

(603.83 KB 2304x2304 catbox_670tj2.jpg)

(589.84 KB 2304x2304 catbox_tcw4yv.jpg)

here's two comparisons of the dpm++ 2m karras fix, before first after second
>>14271 that's what I noticed even in the examples shown in the github discussion. I like a lot of color in my gens so I'm reluctant to use it because of that
(2.12 MB 1024x1536 catbox_3d7uus.png)

seems good to me
I noticed the slightly less vibrant look but KLF8 WD 1.4 vae is already super vibrant so i'm ok with it
>>14271 >>14274 Color should be corrected in post, could try this extension which modifies the final output: https://github.com/Zuellni/stable-diffusion-webui-image-filters
>>14277 eh, to be honest, I wouldn't know how to fuck with with that. I don't ever do any post processing on my images. not even the simple things usually. probably should start though, because I often feel like I've got lower quality than most.
>>14252 I am retarded should've read the guide >Lowering LE learning rate seems to have benefits in separating objects. If you get unwanted objects in your generations, you may want to lower it.
>>14278 >scripts: sd upscale >1.5 >4x ultrasharp >0.35 denoise >multiply the width of the input image by 1.5x, plug that in as the width >scale height to whatever your card can do without OOM goal is for it to generate the image as a series of 1 x whatever images, horizontal bands that have the full width of the final image, because this preserves bilateral symmetry better ie eyes look the same
>>14280 this method might be archaic at this point, i see a lot of people using ultimate sd upscale, but this is built in and it works well enough for me
(2.75 MB 1536x1536 catbox_vn14wz.png)

(2.93 MB 1536x1536 catbox_io1u90.png)

(2.64 MB 1536x1536 catbox_wthk8s.png)

>>14282 fairies are built for rape correction
I noticed -e4 seems to not fry shit with Alpha=DIm LORAs using Adamw8bit? That's weird in my old tests it wasn't that way but ever since getting the hang of d-adaption I have a feeling keep alpha=dim seems like the way to go or maybe it's just the current LORA I'm working on to be honest
>>14284 LR alone doesn't mean much, total steps also matter. Also, did anyone try playing with the weight decay (and maybe beta and beta2) ? I know it was a pretty useful parameter in regular ML shit training but I've never seen anyone do anything but default
>>14281 isn't there that tiled diffusion thing? yeah, this thing https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111 pretty sure it does the same thing give or take as sd upscale. that being said, thanks! do you usually do that after fixing other things like hands and eyes?
(2.25 MB 1280x1664 catbox_l6ibam.png)

(2.75 MB 1536x1536 catbox_9fbhwu.png)

>>14283 can you rape the willing? checkmate!
>>14286 if it requires much fixing i just generate more tbh. faces and eyes can usually be cleaned up quickly and easily with inpainting, bad hands are usually too mutilated to be worth the time. as for order, it really seems to depend on the model and loras, i've gotten consistently good and consistently bad results either way.
>>14289 got it, thanks for the run down, I'll start trying to spruce up my gens.
>>14214 anyone here know how to do live2D rigging? I'm unironically thinking of making a personal assistant waifu for myself
>>14291 OO that sounds pretty cool. How do u plan on making it? using a chatbot or smth?
>>14292 finetuning llama-8B on VNs(maybe?), then using a prompt to make it output a response with a corresponding emotion. I can then use emotion to control the Live2D animation. I think all the pieces are there. I just need to find a good text2voice program and link them all together.
>>14293 Thats fcking cool anon. I wish I could do that sadly I cant code for shit
>>14293 If you succeed make a guide for dumb anons like me got good hardware but zero knowledge about coding
This is reminding me of the one group that tried to crowdfund an AI generated VN but got cucked by Kickstarter
(1.45 MB 1024x1536 catbox_bvvcgh.png)

>>14293 >get hyped by idea >generate random apose girl >stable diffusion laughs at me it's like the program knows what I want and taunts me for being a codelet where is gpt5 which does it for me!
>Danbooru is back up yes, now I can use their tag wiki again Someone needs to scrape their pages for offline use.
(2.02 MB 1024x1536 catbox_ofqlfa.png)

(1.91 MB 1024x1536 catbox_ixagmm.png)

Please touch the belly.
>>14298 >Someone needs to scrape their pages for offline use. Danbooru provides a dump for their whole DB here https://console.cloud.google.com/bigquery?project=danbooru1&pli=1
(1.73 MB 1152x1152 catbox_au2fak.png)

(1.92 MB 1152x1152 catbox_02c9ev.png)

(2.42 MB 1152x1152 catbox_24dpw8.png)

(1.96 MB 1152x1152 catbox_7su4iu.png)

i've made a lycoris version of my shinobu/kiss-shot lora which I think is actually a substantial improvement, the LoRA felt underbaked and this one seems to work better with style LoRAs https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/qkgTxDiD
>>14301 That’s the thing I noticed with LOCONs I wonder how they will change once easy anon implements the new method that Kohya is making them to their scripts
(1.96 MB 1536x1024 catbox_q99tf1.png)

>>14301 Lovely
>>14302 about that, I decided to put the high quality popup on the back burner because it's fighting me a ton, so I'll have a less clean and nice version for the time being until I can finish that popup. This one should still be pretty easy to follow. In other news, LyCORIS implemented yet another algo, I added it, but I don't intend to test it beyond making sure it actually works.
(1.93 MB 1024x1536 catbox_hsqzuz.png)

(1.85 MB 1024x1536 catbox_1qptsm.png)

(1.89 MB 1024x1536 catbox_kzbja5.png)

(1.97 MB 1024x1536 catbox_8ydr2k.png)

>>14213 how did a niggerbrain figure out the b64 code kek
>>14255 i don't get why you don't just fuck with the power limit slider instead, it's so much easier and faster
>>14288 sex with cirno SO MUCH sex with cirno
she's gonna punch a hole through your walls and steal your books
>>14311 you can't just post these and not catbox...
(2.05 MB 1024x1536 catbox_kpfqyx.png)

(1.92 MB 1024x1536 catbox_vt7i5q.png)

(2.04 MB 1024x1536 catbox_2kma0w.png)

running old prompts on new models
>>14068 >apply lora weight to only specific layers I don't think I have the courage to delve that deep but thanks
>>14312 sorry for the delay anon i ate an entire can of anchovies this morning and i just got back from dealing with the consequences https://files.catbox.moe/n3nq6k.png https://files.catbox.moe/co4rz4.png https://files.catbox.moe/1h6odn.png
(4.13 MB 4480x922 tmp9ing11ny.png)

>>14066 nice, time to mass produce sasha in cassock
>>14312 >>14315 they're nothing special btw
If Easy anon is here I want to know if its possible with your scripts if I can switch the GPU I train with to my 2nd one, I did a test and it's less likely to go into high temps when under load (3 slot distance btw so I didn't have the 4 slot distance that could have helped a bit with temps)
>>14318 not him but it's a few cm at best, don't think it'd have made a tangible improvement just power limit the first card a bit
>>14318 it's possible I believe, you'd have to do the accelerate config thing. >>14319 that is true though, an undervolt will likely take you further. also update on the scripts, I'm almost done with the big update, just finishing up the readme.
>>14319 Yeah I'm just trying to make sure I'm within 70 at least and don't go into the high 70s or 80s, btw would a top AIO help with temps? I wanted to upgrade my CPU fan but I was wondering if a top mounted radiator would help with air flow
why do style loras look drastically different on different models?
hey all, This update kinda sucked to make. As it turns out, thinking about the logistics of 125 values being packed into 5 variables is tough to make work with popups. Especially when the popup I was gonna make got totally derailed. I'll get that running eventually. for now... Here you all go, new update to the scripts. changelog: - Updated sd-scripts and LyCORIS - added all of the new huggingface commands to the `ArgsList.py` that was added recently by sd-scripts, but I didn't include them in the popups. - Added support for the new algo LyCORIS created, lokr, in the popups. - Added support for the new block weight training that kohya implemented. There is a lot to it, so I suggest you read the documentation [here](https://github.com/kohya-ss/sd-scripts#4-apr-2023-202344-release-060). - Added a mode to reading a json that allows you to specify the folders still. - XTI implementation will be in the next update links: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/releases/tag/installers-v6 Thanks for using my scripts people, I enjoy being able to see people make use of my stuff. at the very least, I managed to get this out on my birthday, so good enough to me, now if you excuse me, I'm gonna crash into my bed, as it's currently 1AM and I need to be awake for work in the morning, as well as birthday stuff my family is sure to throw at me in like 8 hours.
>>14323 Happy birthday anon and thanks for all your hard work, once I can get both of my GPUs at autism safe temps for training I'll be sure to test out all the new stuff
>>14324 Hope you have fun with the block weight stuff, I really struggled in making it as good as I could. I went through so many designs of the popup that didn't even get finished in the end, lel. but either way, I think my current version works well, so I'm happy.
>>14325 Yeah for now ever since the whole new methods implemented I just went back to the simple nature of just baking a normal LORAs but hopefully this whole autism with block weights will help with making trained stuff a lot more better (but honestly I think autistic taggers and data collectors will still make best LORAs kek)
>>14326 since, locon came out, I've been preferring to train them, at least for styles, vs lora, as I usually see better results with smaller file sizes. I mean, my Tab_Head locon is dims 4/8 and was like 90% of the original dim128 I made, and tenroy was a solid 95% equal to the original artist (despite his inconsistency). I really should think about using locon for characters as well, just to see what I can do.
(2.54 MB 1536x1536 catbox_t6dyni.png)

>>14216 seems interesting. How exactly do you load Locons anyway? is it exactly like Lora?
>>14330 I don't know if you can load it natively now but I used https://github.com/KohakuBlueleaf/a1111-sd-webui-locon as an extension
(1.70 MB 1024x1536 catbox_5llkqe.png)

(1.78 MB 1024x1536 catbox_g6ndlc.png)

(1.81 MB 1024x1536 catbox_3kbd1e.png)

(1.71 MB 1024x1536 catbox_273red.png)

(1.69 MB 1024x1536 catbox_ir7sfp.png)

(1.78 MB 1024x1536 catbox_qz4j68.png)

(1.76 MB 1024x1536 catbox_g4wyt1.png)

(1.63 MB 1024x1536 catbox_kzat85.png)

>>14332 I dig this style
>>14279 The recook with this looks very promising. Just don't be schizo with negs, as dog penis in negative makes it do a slightly dog-like penis...
(716.89 KB 864x864 06747-2214644397.png)

>when no one uses your loras
>>14336 not a fan of dog dick but pic 2 looks pretty good, I'll be sure to give it a go with futa
>>14337 Thanks, it's still a bit gacha-ey but it's superior to the previous versions. Couldn't get all the nuance of foreskin unfortunately, maybe an inpainting lora could be made for that? idk
(2.01 MB 1024x1536 catbox_ukvvzs.png)

(2.08 MB 1024x1536 catbox_s0si19.png)

(2.07 MB 1024x1536 catbox_bfszlx.png)

(2.04 MB 1024x1536 catbox_vbheap.png)

>>14335 Don't cry anon, I probably have your loras downloaded. I just have a huge lora backlog and I'm lazy, also this flat cutie is gonna drain me completely... I'm sorry my wife..
>>14335 Which are your LoRAs?
Inpaint/Img2img Upscalers I was doing a touch up on an image with inpaint but then when I upscaled it in Img2Img, I lost a significant amount of detail in the areas I inpainted, is there a correct order a settings when doing this?
(1.13 MB 768x1152 style-seraziel-v5.4_NAI.png)

>>14338 Seraziel LoRA v5.4 uploaded, also updated cover image and weight/model comparisons. https://mega.nz/folder/53gW3bbK#NxXyJ0l9T3JyAXolJfpBDw
>>14342 based, thanks for your hard work
>>14342 >v5.4 I respect your autism anon, maybe I'll use this to make some futa foxes
>leaked 5090 news >32gb GDDR7 holy kek, yeah I guess I'll wait to spend money on a new gpu for that if ISREAL
>>14328 10/10 Drill hair
(1.84 MB 1280x1664 catbox_ynepam.png)

(6.93 MB 2960x3728 catbox_k295fo.png)

(2.08 MB 1280x1664 catbox_ak59mr.png)

(1.43 MB 1024x1488 catbox_bod3mi.png)

Trained BigrBear into a Style LoCon.
>>14335 My LoRA folder is up to 50+ GBs worth of LoRAs that I hardly use, I'm just hoarding them.
>>14347 Holy based, I was thinking of training this myself. Thanks anon.
>>14330 medusa lora pretty please?
>>14336 #4 is fucking perfect, 10/10
>>14342 >he delivered again x5 i'm sorry i made you go through all this kek, thanks a ton though
>>14349 Thanks. its currently doing another run with a higher resolution. ill drop you a link after that anon.
>>14343 >>14344 >>14352 Thanks. At least I could get a satisfactory result with current lora tech. I don't think I can really push it further for now without making it higher than dim 128 or locon(/loha?) but I'd still be on it next year with my 2060. (And I mostly don't know the limits of it, so idk if I hit theoretical limits or there are things I could improve) >>14351 Dog cock takes a bit of gacha for txt2img but it happens to work really well sometimes, and even though I haven't tried yet, inpainting probably works okay. Next up is Frip/Frippy/Fripsite. Currently combring through their fanbox with kemono to retrieve uncensored versions or do them myself with PSDs. Anyone here would happened to be subbed to them? There are recent posts missing (maybe some older ones since I don't see PSDs for some) Take some clothed holos for the road
>>14354 Could you catbox some of the canine penis ones? I'm having a hard time getting it to work with my settings.
>>14354 (along with dog cock weirdness)
>>14354 >>14355 seconded, would love those
>>14354 >Dog cock takes a bit of gacha for txt2img but it happens to work really well sometimes, and even though I haven't tried yet, inpainting probably works okay. I've seen it work with human holo and that's more than enough for me, thanks again also HOLYSHIT THE FOURTH ONE
(969.69 KB 768x1152 catbox_94jilg.png)

(985.48 KB 768x1152 catbox_q6xmbk.png)

(965.77 KB 768x1152 catbox_wezgxx.png)

(944.03 KB 768x1152 catbox_q7rd95.png)

>>14355 There are examples on mega that have parameters (if mega doesn't strip), but here you go, 3 dog cocks, one human. May be repeats from before since I didn't generate too much. >>14358 Yep. Though I don't know how much gacha is acceptable, though for futa most images are garbage in general anyway. I don't have much good futa images (for my standards, otherwise more than 50% is acceptable for h/d/g lmao) out of the about 100 I generated for testing. I'm also not a very good proompter.
(999.30 KB 768x1152 not how knots work.png)

>>14359 can i trouble you to catbox this one too?
(1000.61 KB 768x1152 catbox_pumhn8.png)

>>14359 I don't think I used it for any of these, but you can try to use Latent Local UpscaLer (LLUL). It's straightforward in txt2img (never tried in img2img), and can help generate better eyes / genitalia. Don't go too hard on it though. Also I've never touched the extra parameters. Drawbacks are the shape and size of the zone is very restrictive. >>14360 here ya go same generic prompt though, maybe someone more devoted to dog cocks can work out a better proompt for it
>>14361 >maybe someone more devoted to dog cocks can work out a better proompt for it NOW THIS LOOKS LIKE A JOB FOR ME thanks again btw
maybe a weird question but can somebody explain to me why something like futa or even dog cocks are something people wanna see and fap to? I don't wanna kink shame anyone but I just cannot understand
(1.54 MB 960x1152 catbox_6llk59.png)

(2.05 MB 960x1440 catbox_4gehu0.png)

(1.77 MB 960x1440 catbox_mcuge6.png)

>>14363 idk man they make my pp big I don't question it also second image is cursed click at your own risk
>>14363 >but can somebody explain to me why No
(1005.01 KB 768x1152 catbox_n4nd4a.png)

(987.79 KB 768x1152 catbox_zcpq1o.png)

(976.83 KB 768x1152 catbox_6pl275.png)

(1.24 MB 1152x1536 catbox_mwf07f.png)

>>14363 I don't really know why, but cocks are appealing while general masculine features still remain disgusting to fap to, so I like futa. Also a foreskin chad (hence both my futa loras being seraziel and necrosmos). I'm not really into dog cock but it normalizes a bit when you look a lot to furry or furry-adjacent art, but again only on futa. IRL it's just disgusting. Also here are some miraculous gens from previous versions, and an example for the necrosmos one (and inori) that I quickly reran (with the old versions still) because I was testing at low res at that time
What was the correct way to use BREAK again?
>>14363 like >>14364 said they make my pp the BIG pp, especially the latter/both together >>14366 stealing the third one
>>14367 To start your next set of 75 tokens before actually using all of them so your token weights become stronger. >ex: Describe backdrop in one set, BREAK, describe subject in next set
>>14369 NTA but does BREAK make a difference on single block/line prompts?
(1.46 MB 1152x1536 catbox_tdd8o4.png)

(1.41 MB 1152x1536 catbox_usergf.png)

>>14368 big balls are indeed good >>14370 yes, if you're bringing it up because of my catboxes, I just put newlines to organize it better, I don't think it has any impact to newline Take other old gens upscaled. Second one was remastered with the v5.4 instead of v1 lmao (which was more like v3), so signature survived
>>14371 i meant more like <entire prompt here> BREAK and not like <prompt1> BREAK <prompt2> BREAK <prompt3> BREAK would it make a difference in the first case?
So I've only seen one anon on 4chan using the LR Blockweighting that was implemented with SD-scripts recently >LR block weighting seems pretty cool, I was able to bump up the number of screenshots in my dataset and it actually improved image quality quite a bit now if only i knew why it was generating random shit all over the backgrounds has any anon here tested it out?
>>14372 Oh, no, since BREAK resets the token batch so it's a separator, it's not an operator you apply after
>>14374 got it, thanks what happens if you go over the 75 token "limit"? never really understood why there's a limit and why it seems so low compared to NAI's 225
>>14374 does the comma placement matters? I always used it like this prompt1, BREAK prompt2
>>14376 NTA but i swear every time I don't end a prompt with a comma AND a space it just looks worse than just a comma/space/neither (yes i went back and tested on the same seed for dozens of pics)
>>14375 I'm really not an expert in this but I think it's basically BREAK. The limit is probably just intrinsic to the way the text encoder works. >>14376 Basically this >>14377, If I notice some weird blending I know I fucked up a comma, which is why it's so funny seeing proompters that use it without commas
>>14378 >The limit is probably just intrinsic to the way the text encoder works. that makes sense yeah >>14378 if something looks ever so slightly off i usually go back on the same seed and most of the time removing the final space/comma/both fixes it
>>14335 when someone uses 2 of your loras in a single gen
>>14332 it's pretty nice. do you have a link to the lora?
>>14375 as far as I know, NAI's 225 limit is actually 75 * 3, much like our webui is.
>>14375 >what happens if you go over the 75 token "limit"? it does some sort of bucketed blending to keep weights sort of consistent across the length, but i've had outputs change dramatically from crossing the 75 token marker even if it's just a few tokens over, and despite all the denials and paper-posting i remain 100% convinced that tokens at the front of the prompt get weighted more heavily.
(2.23 MB 1711x779 Lora grid.PNG)

My lora preview grid always brings me a chuckle. All my gens seem to have one thing in common
>>14383 >despite all the denials and paper-posting i remain 100% convinced that tokens at the front of the prompt get weighted more heavily I dunno what papers you're even talking about but I can back you up with tokens at the front of the prompt being weighted more heavily when the token limit it crossed in the webui since that's exactly what the PR does. https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/2138#issuecomment-1272825515
(3.02 MB 1280x1920 catbox_sikjm7.png)

(2.74 MB 1280x1920 catbox_zbusd6.png)

(121.98 KB 600x840 3385634.jpeg)


Playing with training setings again since I wasn't supper happy with the results of 8 dim lora. This I'm making my own pop up parade lora and trying locon and lion again. Cleaning up the dataset was pretty simple - de-jpeg, de-sharpen twice, then de-edge.
>>14387 >pretty simple but very long
>>14387 i was extremely excited about the figma/pop up/nendoroid loras but i haven't used them yet
(2.39 MB 1280x1920 catbox_kqqeo0.png)

(2.90 MB 1280x1920 catbox_ixenpi.png)

>>14388 The images are small so it only took a few minutes to run through the full dataset each time. That said I'm also working on a jcm2 lora and left it overnight to de-jpeg which took about 7 hours
>civitai being DDoSd Fucking trolls, I NEED MY OVERBAKED LORAS YOU FUCKS
>>14387 >de-jpeg, de-sharpen twice, then de-edge what is this?
>>14391 downloads seem to work fine, the site's state has just been frozen since this morning
>>14393 Guess I'll keep trying then I get Cloudflare warning pages wherever I refresh or try to dl a LoRA
(2.47 MB 1536x1536 catbox_qfq0nx.png)

(2.33 MB 1536x1536 catbox_vzdo30.png)

(2.57 MB 1536x1536 catbox_9umlqe.png)

>>14381 https://anonfiles.com/q3ydXcj2z1/chii_tsumami_tsumamare_V2_safetensors only problem is that it sometimes refuses to gen nipples without prompting for it and flat chests are a bit difficult but other than that I like it
>>14395 nice, thank you
(2.86 MB 1920x1280 catbox_z2vbl9.png)

>>14392 Models for image enhancement/upscaling https://upscale.wiki/wiki/Model_Database Been using them lately for cleaning up datasets
(2.22 MB 1152x1728 catbox_xkz99s.png)

>stumbled back upon tsumetoro/natsupoi I'm starting to hate AI and LoRAs, the list of artists to do is way too long I haven't even finished building the frip dataset
(2.59 MB 1536x1536 catbox_oy5rhj.png)

(2.28 MB 1536x1536 catbox_bxp11j.png)

(2.42 MB 1536x1536 catbox_daxnaq.png)

(2.37 MB 1536x1536 catbox_yef2nk.png)

hm... a round belly is fine too And milk is the base of our evolution! It cannot be bad for us, right?
>>14398 should try your hands at a bosshi lora..
(1.60 MB 1024x1536 catbox_u0i75q.png)

it's so cute, but the hands are so shit
>>14400 there are now 50 loras I on my todo list I'm not even kidding
(816.20 KB 640x960 catbox_4dzqzk.png)

(822.29 KB 640x960 catbox_d4bylt.png)

(842.67 KB 640x960 catbox_0rbmlo.png)

(898.58 KB 640x960 catbox_qd9oh5.png)

The ATF lora guy reuploaded all their loras btw. Seems like a lot of good artist but haven't been able to test them yet >>14400 >>14402 Here's a bosshi lora by said guy, reupped to catbox just in case https://files.catbox.moe/qrptzi.safetensors
>>14402 consider the following:
>>14404 actually nvm, helen isn't in that pic so 0/10
>>14403 shiro my beloved I'd post some previous gens with the wagashi lora (criminally underused) but I only did low-res shit at the time Also waiting for tests with bosshi lora to see if it's good >>14405 Fellow foreskin chad ellen is pure sex, needs (saggy) balls though
>>14406 >ellen is pure sex 100% >needs (saggy) balls though just mix in some zheng kek
>>14407 >zheng Zheng was a mistake. The balls seem paper-thin and way too droopy. Could work for gens though.
>>14408 yeah that's why i said mix in SOME though maybe it'd work better if done as a concept lora now that i think about it
>>14409 Did anyone try to make concept loras by simply cropping pictures to select the feature? Would that just make it an "inpainting lora"? If I didn't have that much artists to work on I'd do that with the seraziel penises, but pioneering that might be a bit too much
>>14410 idk if that'd work, ideally we shouldn't have to do that but idk if we can get perfect dicks with the current lora/locon/loha/whatfuckingever technology now back to my knots...
>>14412 what style is this?
>>14413 Nora Higuma, on gayshit repo
(1.92 MB 1024x1536 catbox_kxaur4.png)

(1.96 MB 1024x1536 catbox_wlkk7s.png)

(1.94 MB 1024x1536 catbox_mk0e8x.png)

(2.20 MB 1024x1536 catbox_p1eb7q.png)

>>14412 Great stuff and I need to use this LoRA more, I love Nora Higuma.
(1.40 MB 1024x1536 catbox_9gznw5.png)

(1.72 MB 1024x1536 catbox_nle778.png)

(1.66 MB 1024x1536 catbox_3iqa5p.png)

(1.69 MB 1024x1536 catbox_n6uzso.png)

I meant to share this a while ago, but I got caught up in working on the easy training scripts that I kinda forgot. well anyways, here are some preview gens of zerogura, I'm not posting this one up on civitai so you can only get it from my mega this time. Hope people find some enjoyment out of this. Warning for people that do use it, about 1/3 the dataset was pissing so it's very possible with this lora, just a warning ahead of time, and no, I didn't make this style for the pissing. mega: https://mega.nz/folder/CEwWDADZ#qzQPU8zj7Bf_j3sp_UeiqQ/folder/fFZG0BjJ
which versions of the pop up parade/figma/nendoroid loras are the best? don't really want to test each one and it's not too clear from the grids either
(2.38 MB 1024x1536 catbox_1x7a1h.png)

(2.10 MB 1024x1536 catbox_zxtrpk.png)

(2.29 MB 1024x1536 catbox_itgrcf.png)

(1.95 MB 1024x1536 catbox_rdzegs.png)

>>14416 Yeah, I definitely didn't download this style for the pissing.
>>14418 fair enough, I'd like to just consider the pissing a bonus of sorts, people who like it totally can use it for that, after all, it is good at it.
every single fucking time i restart the webui for whatever reason it gives me this error and i have to restart SD itself <RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument mat2 in method wrapper_mm)>
(2.25 MB 1280x1920 catbox_7g6a5m.png)

(2.22 MB 1280x1920 catbox_i900do.png)

Finally a chibiusa lora
i'm a literal birdbrain, i only JUST tried img2img after all this time was about to ask why it isn't deterministic but then i noticed that you can use a seed too
(5.78 MB 2304x1536 catbox_w2nfvb.png)

Scary stuff
>>14423 nice
(1.60 MB 896x1664 catbox_uxmlvq.png)

Considering a GPU upgrade from a 3060 ti. would a 6950x + Linux be a viable option? or just stick with Windows and get a used 3090?
Forgive me for the off topic shit, but >nahida lora >nahida prompt for LLMs >Whisper speech to text We actually live in a world where we can bring 2D into 3D. What a time to be alive
>>14426 >going from nvidia+windows to AMD+linux you're gonna have a rough time
(5.01 MB 2304x3456 catbox_dpi3w1.png)

>>14359 I hate gimp
(2.07 MB 1024x1536 catbox_g02zt5.png)

(2.03 MB 1024x1536 catbox_k2q7t0.png)

(2.07 MB 1024x1536 catbox_axxcmt.png)

(2.13 MB 1024x1536 catbox_fn5fj6.png)

Made for corrections.
>>14426 Go the used 3090 right. AMD cards are cucked. You can still go Linux.
>>14421 Where did you get it?
>>14427 we could probably get a text thread going if there's enough interest, LLMs are already exploding and there'll be another big boom in character chatbots (and therefore character ERP) with the WD+pygmalion model that's coming up
>>14433 If someone can make a thread with what we need to set up and run everything I'm down. I saw that some of these models need 24GB VRAM so I can play a role in testing.
>>14433 >>14434 there's a thread on prolikewoah /hgg/ but honestly things are moving too fast and boards like this have too little traffic for a thread here to be worth much for anything other than news unless it's being run by someone involved in pushing the frontier. current state of things is basically how they were for imagegen a couple of weeks after the NAI hack, ie crude finetunes and incomplete implementations. they're going a lot slower because textgen just requires more hardware so there's fewer people involved.
>>14380 You mean one ;)
>>14416 Who the hell is zerogura? That artist is called ama mitsuki
(1.33 MB 768x1408 catbox_x4jz9s.png)

(1.40 MB 768x1408 catbox_2khjp6.png)

>>14429 still hate gimp though very useful to make edits can you tell I like big balls?
>>14439 balls need to be at least jawbreaker-sized and knots need to be at least twice that
(210.03 KB 699x1129 Screenshot_20230409_134311.jpg)

which of these should i train today (/try/ to train if it's upshorts since that will be hard to wrangle)
>>14443 sindoll
>>14439 >making big saggy balls is as simple as free select, scale and quick brush and blend >making cum is as simple as white brush with a bit of transparency I know inside of me that I should probably have a catalogue of cocks and other props by now since it works amazingly well
(1.11 MB 768x1152 catbox_6grrw1.png)

>>14445 cum example may try to outpaint for holo
>>14443 aruurara
>>14386 Yeah, the language model works from left to right, changing a word in the middle has no effect on what it outputs for the preceding text. And it's trained to summarize the caption's meaning into the last token's output, as only that gets matched up with images in its training. SD uses the full-length output to squeeze more information out, so it sees the meaning of the start of the prompt repeated dozens of times. Crossing the 75 token marker means giving the language model two separate inputs, so unless you're schizoprooompting word salads that could be cut anywhere and still mean the same thing, you're gonna get drastically different results.
(1.52 MB 1024x1536 catbox_bmhtmg.png)

adventures in making the most retardedly lurid lora mix i can find
>>14443 Sindoll or the top one.
>>14443 sindoll and totally not his alt cervina7
is snuffy a tranny or some shitlib or something? wanted to train a lora, raccoon girls are cute
>>14451 i think someone on /trash/ already did cervina7 (someone might have done sindoll too but a lot of /trash/ LoRAs are horribly outdated and poor quality) I'm finishing a different kemono LoRA I forgot I had the dataset for then I'll take a crack at sindoll
>>14453 >>>>/trash/ fuck I forgot that place exists.
>>14453 neat, who are you doing? also yeah the trash ones are... trash. wow what a surprise
>>14455 they at least try go look at /d/, they're all using that one dogshit civitai futa lora or spamming fried gape
>>14455 I just trained crayon (yes their name is literally just crayon): https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/S8plGRLB >>14454 Yeah, I'm a furfag so I occasionally check to see how their /sdg/ is doing. Surprisingly, a few really cool non-furry things have come out of there like the OoT style LoRA. Unfortunately they don't have a lot of people doing training and most of their stuff was made around when LoRAs first released so they're using poorly optimized settings. Their thread is still interesting because it feels like its own little bubble separate from the rest of the 4chan SD threads, I'd recommend peeping their OP just because they have (furry and non-furry) resources that aren't listed anywhere else.
>>14457 yeah i know crayon, looks a lot like what crosskemono E/F can do
>>14456 they try but just like the other anon said the most successful one they made is the OoT lora ironically
>get aio >had shitty screws >temps are the same as my stock wraith cooler yeah I'm just going to return this one and get the arctic ii
I saw people running their own language models? I'm willing to download the leaked facebook lama models if we can actually run it like the waifu chatbots and someone makes a retard guide on rentry
>>14462 https://github.com/oobabooga/one-click-installers you can indeed run it like a waifu chatbot and setting up the software is a one-click installer, once you're done you can go to https://booru.plus/+pygmalion to find personalities or use them as templates to write your own
>>14463 thanks, getting closer to having an AI loli anime gf
>>14433 >WD+pygmalion model that's coming up Really hope this turns out to be good because nothing still compares to Turbo yet. Concerned if something like GPT4 is ever even going to be possible to run locally.
>>14465 better to wait if Ada titan comes out or if the 5090 32gb model is real to even see if it's feasible to run something like GPT4
>>14466 You couldn't even fit Turbo on that.
>>14467 pain, huge language models are the last frontier for AIbros huh
sindoll added, next I think I will start work on fundoshi/sarashi since prompting them is fucking impossible https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/K443SBhL
>>14098 What lora are you using for the style?
>>14471 looks like cutesexyrobutts
(1.65 MB 1024x1536 catbox_goidof.png)

(1.73 MB 1024x1536 catbox_wbhhg7.png)

>>14470 Nice and comfy.
>>14473 Your gens are giving me a HARD time anon. now gen more pls
>>14474 K have a mix of new stuff and something from yesterday, now I sleep.
>>14469 Can this only generate anthros? Should I use a furry model with this?
>>14476 It's only trained on anthros so I didn't test humans, but you can give it a shot and see what happens. Those gens are on based64 because it was trained with Danbooru tags but I'm doing a crosskemono test right now which will use e621 tags.
(1.99 MB 1024x1920 catbox_gs1f05.png)

(2.31 MB 1024x1920 catbox_znis58.png)

(2.46 MB 1024x1920 catbox_70had4.png)

>>14446 took more work than expected but eh
>>14477 well, training a lycoris on crosskemomo with the same settings as nai came out completely deep fried above 0.1 weight so now I'm going to try baking a regular LoRA with adamw8bit instead of dadpt
>>14479 oh it's because I'm retarded and though crosskemono is a finetune when it's actually a weighted block merge of multiple checkpoints and loras, some of which use danbooru tags and some of which use e621 tags what the fuck
>>14479 >lycoris Is that locon? How does it fare against loras?
>>14481 nta but LyCORIS is a blanket term for LoCon, LoHa, ia3, and Lokr.
>>14481 it's the same theory as lora but trains more layers of the model, should theoretically provide better results than loras but it's honestly hard to tell
>4chin still thinks 768 is a placebo >meanwhile me and a few anons training over 1000 pixels kek
(2.53 MB 1024x1536 00725-4127076101.png)

(2.56 MB 1024x1536 00732-184151502.png)

(2.36 MB 1024x1536 00780-2762166044.png)

(2.48 MB 1024x1536 00788-298191212.png)

Happy easter
>>14482 Do I need to install something to use lycoris?
>>14487 yes, https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris it only allows you to use it in the built in version though
>>14485 I just saw the post you were talking about lol
(1.32 MB 864x1152 catbox_flltq6.png)

(1.17 MB 864x1152 catbox_azckm6.png)

(1.51 MB 864x1152 catbox_4hmyed.png)

(1.44 MB 864x1152 catbox_en38o0.png)

sindoll LoRA, "de-furrying" an artstyle is really interesting, makes me wonder how close it'd be to the artist's actual works if they drew humans
(2.46 MB 1024x1536 catbox_cna5to.png)

(2.64 MB 1024x1536 catbox_vqxorz.png)

(2.47 MB 1024x1536 catbox_pp3zp4.png)

(2.41 MB 1024x1536 catbox_voq1mu.png)

>>14486 So he is alive
>>14491 What's with that model anon? Is that your personal mix?
(1.58 MB 960x1152 catbox_qeqv3v.png)

(1.84 MB 960x1152 catbox_nw0ulz.png)

(1.71 MB 960x1152 catbox_1mz1a1.png)

(1.78 MB 960x1152 catbox_yl40fh.png)

(3.05 MB 1456x2048 catbox_ghei5c.png)

(2.67 MB 1456x2048 catbox_mny86e.png)

(2.34 MB 1456x2048 catbox_xp0pmd.png)

(2.73 MB 1456x2048 catbox_im4743.png)

BigrBear style is now finished. to the anon who was asking for the other day here you go. https://civitai.com/models/35278?modelVersionId=41557
(2.96 MB 1920x2496 catbox_iyc52x.png)

(1.48 MB 1360x2048 catbox_nfayon.png)

>>14485 >try 1024 >halve batch size >training the seraziel lora would take a full day on my poor 2060 oh god make it stop
>>14491 Super cute bunnies
>>14497 try getting lucky like me and finding a used 3090 for a great deal that still has RMA
>>14461 get the 280/360 also yeah no reason to ever get anything but the liquid freezer 2 kek
>>14499 1 year left on mine but the seller can't remember where/when he bought it so i can't register it on the evga site kek i'm fucked
>>14499 I'm on a laptop unfortunately
>>14502 >seraziel anon has been doing this on a 2060 mobile we will build you a statue
>>14471 >>14472 yep, CSRB lora elevates my lazy basic bitch proompting using remacri so it doesn't look as good as latent/anime6b (picrel are two of my gens upscaled again with remacri and i prefer the original anime6b/latent ones tbh) also i just noticed that yukari's eyes turned purple-ish again, sd keeps giving her purple eyes (no matter the emphasis on brown eyes) thinking it's the vocaloid instead, very annoying
she's PISSED (that's hot)
>>14486 is there an egg laying lora?
(2.42 MB 1536x1024 catbox_e0k64t.png)

(2.51 MB 1536x1024 catbox_uwpri2.png)

(2.62 MB 1536x1024 catbox_v0291o.png)

(2.71 MB 1536x1024 catbox_m25dy7.png)

>>14492 Personal mix it's using test model as a base but you can get similar output like I have if you use loliDiffusionV0.8.8 based 64 mix
>>14509 i should mess around with the same seed more often
>>14443 waai fu from arknights because i need my tiger wife and the existing loras aren't that good
https://civitai.com/models/34871/change-a-character-brainwash-your-waifu-today Was browsing civitai while bored and found a new mind control/brainwash Lora, anyone want to try it out? I'm too busy to do so myself but thought I'd share
Haven't touched AI stuff for a pretty long while after prompting extremely fucked up shit but I ended up with like 3 gbs of LORAs so guess I'll test out all this shit and post results here Civitai has a surprising amount of loli stuff despite their rules. Too many uploaders and not enough unpaid trannies to delete all the underage anime characters.
>>14513 >Too many uploaders and not enough unpaid trannies to delete all the underage anime characters. still wondering why no one's set up a botnet to flood 4chan with lolis, would be so fun to watch the unpaid trannies screech "NOOOOOOOOOOOOO THIS IS A BLUE BOARD MUH (((ADVERTISEEEEEEEEEEEEEEEEEEEEERS)))"
>gamers finally whining about VRAM numbers thought only AI users like us would care but games are getting too memory hungry now
>>14514 Plenty of people have actually set up spambots, it's just that they're used in malicious ways such as to flood threads or kill threads rather than stuff like loli spam. Guild wars general was literally killed for years because someone set up a spambot that would flood the threads and make them unusable whenever it detected certain keywords in the text, the Blue Archive general has a retard that spams gore/scat/blacked porn in an attempt to kill the site, the visual novel general had a spambot reminiscent of the one used to kill the GW2 general and I just saw someone spam nikocado avocado pictures manually for hours in /v/ before deciding I wanted to coom to AI instead of watching him do that. Dude kept getting whacked by mods but he'd evade to keep spamming the same picture and it Makes me wonder why people are like this and why can't they just discuss videogames in peace like normal people.
>>14516 >>14516 >Blue Archive general has a retard that spams gore/scat/blacked porn in an attempt to kill the site >Makes me wonder why people are like this and why can't they just discuss videogames in peace like normal people sounds like the usual JIDF employees at work tbh
>>14516 >>14517 and mods are fucking useless even if you tell them directly that there's a cucked spammer shitting up a general, they want you to report every single spam post
(2.16 MB 1536x1536 catbox_phh2js.png)

>>14486 I must humbly ask how my best try is that one
(1.88 MB 1024x1536 catbox_ur7a92.png)

(1.78 MB 1024x1536 catbox_594xx5.png)

(1.67 MB 1024x1536 catbox_stmajq.png)

(1.70 MB 1024x1536 catbox_h9bdx4.png)

More style LoRA mixing, I should use the fishine LoRA again and see if I can mix it with some of the new stuff.
>>14518 >Report literal automated spam on a niche game thread >Nothing gets deleted I've just mostly given up on that site. I still use it because it's basically the only place I can get discussion for the gacha games I play but holy shit the moderation of the site is worse than worthless and the posters are all on the spectrum. Dude raids a thread for days/weeks with copypastas and walks away free while you do maybe 5 posts and get nuked from orbit because you made the mistake of posting something a mod didn't approve of in a place with a pocket mod despite not breaking a single rule. Sorry for the rant, I'll just go sleep instead of continuing this because I could rant about how awful gacha threads are and how useless moderation on 4chan is.
>>14521 could rant for hours about how awful it is* The typo bothered me too much.
>>14521 >genshit thread >type out diona's name >get permabanned for loli
>>14503 kek I'm seriously considering just letting it run for a day to train on 1024, but having no experience at that res idk if my current settings would work
Testing out MBW, jesus I can't imagine how much of a pain in the ass it was to put up those popups. Thanks Easy anon >>14524 honestly with learning rates as long as if you don't have an exponent that will fry your LORA or don't go too fast with the LR It'll be ok, at least with AdamW8bit or anything that isn't Lion
>>14525 Will try it then. The v5.4 was running 2e-4/5e-5 for 10 epochs, 3 restarts and about 18k total steps. Though, as I halved batch size (4 to 2), should I change the LRs too?
>>14526 with LRs I test all of them for every batch, so far I liked what 3e- has done, as for text LR? Kek I just kept it at 5e-5
YOU BOUGHT A 3090 CAUSE YOU GOT JEWED BY HIGH REFRESH RATE MEMES AND ITS (((SHILLS))), I BOUGHT A 3090 CAUSE I NEED TO UPSCALE MY WIFE'S AI-GENERATED TITS TO 4K WE ARE NOT THE SAME
>>14528 the 2nd reason is based, only reason to spend close to 2k on a 4090 too
>>14529 it's closer to 2.5-3k here kek i can't wait for 3k reference 5090s
>>14530 I'd be willing to spend that much if it's 32gigs of DDR7 memory, if it's just 24 I'll wait until it drops on the used market. So literally until 2027 or 2028 until I buy a newer Nvidia GPU
>>14527 Yeah, I may try 3e-4 to reduce the number of steps
>>14531 better be 48gb or i ain't even gonna consider it, 32 is too low to justify something you can get with dual 3090s
>>14533 yeah just reminded me that literally two 3090s that I got still saved a few hundreds than a 4090, I literally only care about VRAM so "muh gaming clock speeds" don't really matter to me. The Ada chips are better for AI but having two cards is better than having one, 48gbs would get meme'd on by every PC gamer GPU channel kek
>>14534 >48gbs would get meme'd on by every PC gamer GPU channel kek those memes aren't gonna last long considering the quality of the recent releases and how until ~2017/2018 or so 8gb were considered overkill
>>14535 Wonder what a game that uses up 48gbs would be like, 8k ultra nightmare settings on everything? maybe 10k
>>14536 considering that the 4090s is already starting to dip below 60 on recent releases while maxing out at 4k and/or needing DLSS (lol, lmao) i'm pretty sure 8k isn't gonna be a real thing for a while frankly i don't get 4k memes either unless you're a single neet with a huge 4k tv all to himself
>>14490 trying it myself and it looks great, you get to keep a lot of the soft and plump style
Does anyone know what style lora this is?
>>14539 5 seconds in saucenao, it's not AI https://www.pixiv.net/en/artworks/106848792
>>14541 very nice, I hope one day we'll be able to train the AI to consistently output tails from behind
>>14542 >from behind and from the front too honestly i don't bother with the tail, it doesn't show up 9 times out of 10 when unprompted and when i prompt for it it's so janky that i might as well just fuggedaboutit
>>14506 nta but I know of one, it's kinda shit, might be one they baked themselves, let me see if I can find the one I was thinking of wow, it's shittier than I remember https://civitai.com/models/9401/birthegg whatever, here is link
she doesn't get hedge mazes
(11.81 KB 1152x720 Untitled.png)

>>14525 you don't even know how much I was wracking my brain to make that work. like, I spent so much time just thinking about how to implement it, and then had to actually implement it, which came with it a bunch of error checking I had to make sure was there as to not fuck everything up. and even then, this wasn't the popup setup I wanted to go with. I ran out of time. this was the shitty mockup of the actual popup I wanted. but that ended up being a huge pain in the ass, so I had to shelve it. It will eventually become a thing, eventually, but not right away, other things to work on. so long story short, glad you like it.
>>14546 probably a retarded question but why did you choose popups instead of something like this?
>>14547 natural progression of the program, back when I made it, it was stupid simple, and there was like 10 popups. I've been wanting to make a proper UI, but I have this habit of wanting everything to be really polished, so I keep putting it off
>>14548 that being said, I had definitely made plans to have it be a UI, I was thinking of using PyQt5 or 6 so as to avoid the hell hole that is gradio
>>14548 should really consider doing it at this point, the popup approach sounds more and more unsustainable as more shit gets added
>>14550 yep, this is the tipping point for sure, the block weights actually killed me, so it's time I stop putting it off and make a proper UI
>>14551 >PyQt5 idk why but every single QT ui i've ever tried just breaks on my machine if you end up making the UI you could integrate all the popups as on-hover tooltips instead
>>14552 hmm, to be entirely honest, I've pretty comfortable with tkinter, so I could just make the UI in tkinter. it might be less polished/flashy but I could make it work. and it would also mean people don't need to install more things... decisions, mistakes... in all seriousness though, I have never had an issue with PyQt5 so I can't say I ever thought about compatibility as an issue
>>14553 >in all seriousness though, I have never had an issue with PyQt5 so I can't say I ever thought about compatibility as an issue dwai i'm just memeing though i did notice QT shit hanging an awful lot on my system
>>14554 I've had that same issue with my own popups lel. sometimes it just takes forever for it to launch.
pippa anon... it's time...
>>14504 i'm pretty sure that's just csr always coloring brown shit purple
>>14557 thought so at first but nope, happens without the csrb lora too "yukari" seems to almost guarantee purple eyes which is kinda puzzling, yukari (the vocaloid) isn't that popular yeah there's touhou yukari too but if i prompt just yukari/akiyama yukari it'll give me purple eyes AND hair (thankfully you can hijack the hair really easily) so it's obviously not going for the hag
Anyone tried playing around with weight decay? I noticed one or two of my old loras were done at 0.6 instead of default 0.1 (not dadapt) but idk how that affected them
>>14558 then again could be (probably is) a combination of both yukaris, i think the hag has way more purple eyes pics than yellow eyes pics
(1.64 MB 960x1440 catbox_hvbwn5.png)

(1.74 MB 960x1440 catbox_md88u4.png)

sigh new weird waifu i probably shouldn't stick my dick in just dropped https://youtu.be/feltO0inFfo
>>14561 is there a bandaged chest lora or embed? never works properly when i prompt for it and i refuse to give shion and reimu bras
>>14563 yes I just made it last night >>14493
>>14564 >sarashi i'm literally blind thanks anon
>>14562 i'm not clicking on your vtuber shit
>>14506 This was a self-made lora, I plan to make it better since it requires multiple keywords to be consistent. (egg laying, egg:1.2), spread legs, (lying:1.1), spread pussy were used in that gen, but the prompt was bloated so you probably don't need to weight it that much. Since I don't know when I'll bake it again, here's the beta : https://files.catbox.moe/cccx69.safetensors
(2.13 MB 1536x1024 catbox_xjh9td.png)

(2.32 MB 1536x1024 catbox_z9puyk.png)

(1.96 MB 1536x1024 catbox_4339gz.png)

(2.17 MB 1536x1024 catbox_yq9ayo.png)

also, here's some bonus eggs
>>14568 where can someone find that egg laying lora? asking for a friend
>>14569 anon, literally look up like 2 posts
>>14570 I am an idiot thank you
>>14525 Is that your post on /hdg/ with the huge grid?
(881.90 KB 4000x1803 catbox_u9dsga.jpg)

I guess this is one of those cases where there's no way I can get significant improvement short of discovering a trove of high quality data to use. Nothing I try ends up better than the original 128dim LoRA.
>>14573 I have jerked to that artist's jeanne alter so many times.
>Normie channel doing "muh dark internet stuff" making a compilation of the darkest moments in Vtuber history it's over... I gotta learn language models so I can get away from Vtubers and make my AI anime girlfriend
>>14575 >we could live to be able to generate our own big balled futa AI wolfgirl bros we're back
(1.75 MB 1152x1152 catbox_qevpp9.png)

>>14574 same, beast friends is my favorite h-doujin and she is my favorite h-artist of all time unfortunately she doesn't post erotic art outside of doujins, most of her other art is paywalled, and the only way I think I can improve my dataset is running the manga cleaner model which requires making pixel masks for every page's text bubbles and that's a herculean task that I'd rather not do
>>14577 Have you tried lama-cleaner?
>>14578 I have not, I thought that was mostly for denoising. Can it be used for cleaning manga pages/panels, and is it good at it?
>>14576 But I'm doing that right now?
>>14579 Well, turns out it can though YMMV. 100% when it comes to removing small floating text and watermarks, text bubbles depend on size and position. Black censor bars are also hit or miss but the more I can deobfuscate messy images, the better.
lolifying characters can be fun
any highres training that anons decided on here?
>>14583 what do you mean by this
>>14584 what high pixel resolution did anons decide on using for LORAs training
>>14585 Well, the AI learns based on features in the image rather than pixel space so you should pick the highest resolution your GPU can reasonably handle. Just like every other hyperparameter there's no one size fits all option and it heavily depends on what you're training and how distinct the features are in the images as well.
(1.77 MB 1024x1536 catbox_z7wrht.png)

(1.73 MB 1024x1536 catbox_on9xb5.png)

(1.75 MB 1024x1536 catbox_pepj8q.png)

(1.84 MB 1024x1536 catbox_v59tks.png)

>>14582 Same, anon.
(3.62 MB 2816x2816 balls.png)

>>14580 didn't specify what I meant by generate, but I'd add a LLM behind with live voice chat (and immersive VR) but I am doing it too. trying out regional prompter (I tried to do holo+furry holo standing up, maybe kissing, with cocks touching) ((yes I'm not good at regional prompter, and the time spent doing this using a reject gen didn't help)) >>14582 that's taking it too far WHERE ARE THE TORPEDO TITS AAAAA
>Panty pull lora Grabbed this off civitai. Seems to work but is pretty gacha and it tends to double the panties for some reason sometimes.
>>14592 bro your vae????
>>14593 It is enabled, the loras just like that.
Alright I don't know shit about BNHA but invisible girl is pretty hot.
>>14594 >>14593 oh nevermind, I had dynamic thresholding enabled and didn't notice. Woops.
>>14595 she was hotter when she was invisible tbh
"i'll do a run of ror2 before baking" >nearly 6 hours later jesus christ i need to uninstall it, it's too dangerous
Why does the final uploaded lora name always differ from the lora name makers saved it as?
>>14591 i'm surprised you've never genned myuri
>>14595 i know enough about bnha to know that it's absolute garbage with a few good girls but yet, invisible girl is really cute
Cute peeing elf from princess disconnect
(1.92 MB 1152x1152 catbox_52igra.png)

(1.56 MB 1152x1152 catbox_urixvd.png)

(863.05 KB 768x768 catbox_k0x17f.png)

i'm not sure where these little text boxes are coming from but they're pretty cute sometimes
>>14598 ror2 is based the opposite is happening for me, I can't stop genning futa holos so I barely even game >>14600 I forgot she existed, last time I watched the show was a really long time ago. Problem is idk if it know who she is since seraziel barely has any pics of her, and she's mostly not fully in frame, so... Tried it, first gen went okay, added tags for the frame but second and third were holos (including this loli holo), so weighted the grey tags a bit more and it seems to work okay, just used llul on the face cause eyes were fucked up Not in a uoh mood currently but still
>>14577 our favorite franchise might be a dead rotting corpse reanimated by japakikes and forced to dance for years on end but at least kadokawa made beast friends canon
>>14605 yes they're THAT desperate (though this happened before KemoV)
>>14604 she's holo's daughter, seraziel has a few pics of her but this is the best one not exactly a good lora candidate considering she's silver holo in her teens
What models have the best LORA compatbility? I see most people stay on AOM2 NSFW and avoid AOM3 completely though I don't know why. I've also seen a lot of Anything 4.5/Anything 5 and a model called counterfeit? Any recommendations? I've mostly stuck to only using AOM3 and Based65 so I want to branch out. And did that anon ever finish the final completed mix of Based65,66 or whatever the final name was?
>>14609 b64v3, 7th layer v2(a), maybe v3, aom2 nsfw, aom3a1 and only a1
(2.05 MB 1152x1152 catbox_61rpcx.png)

>>14605 That's hilarious, I tried to play some of KF3 back when I was spending more time studying japanese but I didn't get very far. And honestly, as long as they keep coming out with new Friends the community will keep the true spirit of the franchise intact.
Might make a chinese dress lora/whatever if I figure out how to make a concept lora and if I can force myself to stop hoarding pics at some point
(1.70 MB 960x1344 catbox_ugrkb3.png)

>>14612 i haven't had any issues prompting cheongsam (I even get crazy variants like picrel sometimes), is there a specific variant you're going for?
>>14613 nothing specific, no just want to add some variety and possibly reduce the gacha on crossed legs/one leg up poses
>>14611 there's an english patch now, idk if it's good on one hand i refuse to play it on principle on the other hand i could try to get shoebill on a third hand i'm already doing arknights + genshit on the side so
>>14608 yeah, seems so >>14610 >b64v3 is there a link for that one? I only have b65
>>14609 i use 90% based64v3 and 10% aom2-r34-sf1
>>14576 >ywn have momiji threaten to knot you for tresspassing unless you buy an 80gb commercial grade gpu it's so over
>>14609 mainly based series and anything v5. v5 as usual really hard to gen nsfw though. I wonder what mix that you could use so it has nsfw capability while maintaining the core characteristic. It can fit kl-f8 easily without oversaturated even with typical civitai fried lora too.
>Lora that always worked before doesn't work before >No error log on console or anything Why are computers so hard
>>14623 >doesn't work before my man got a time machine
>>14624 I wish
https://bunkr.ru/a/6vWSsP8Z If the anon that did Myrtle, Suzuran and that stuff is around then I'd like to inform that your bunkr link is down. I have the loras downloaded, just a heads up.
>Look up sample picture in civitai >Maker is using bad image embeds but doesn't tell you where to get them >Is using other loras but doesn't include them >Model isn't listed or in civitai I want to scream
Here's some of what I found in civitai and my thoughts on them https://civitai.com/models/34871/change-a-character-brainwash-your-waifu-today Works half of the time and feels pretty gacha https://civitai.com/models/8275/empty-eyes-lora-utsurome-hypnosis Pretty good. Used to not like but the maker has heavily improved it https://civitai.com/models/32360/corruption https://civitai.com/models/20593/shirt-pull-or-test-sex-act-lora-101 https://civitai.com/models/23440/panty-pull-or-test-sex-act-lora-117 It does exactly what it does on the can, not much I can add. https://civitai.com/models/26818/cum-bath If you mess with it then you can generate a girl drenched in cum. Useful for when you want to cum buckets and the AI just generates drops https://civitai.com/models/26342/monster-or-sex-act-lora-79 Undecided but probably worth a mention since usable sex loras are rare
>>14628 Loras that mainly made by anon and makes it there are pretty good. Other than that, can't say much since it's usually fried despite it being useable without much tinkering to "recommended weight" while mixing with other loras.
I wish there was a embed/lora to remove navel nipples
>>14630 The nobeta one works pretty well I haven't been here in a really long while so I just dug though civitai for new toys to play with and share whatever looked decent. Any suggestions on LORAs to download or fun things people here have made?
(2.29 MB 1536x1152 catbox_hij2yx.png)

>>14632 You can always check gayshit for some fun ideas since I don't know whats fun for you.
>Spent the entire day setting up a VM for my Hydrus and AI work >can be remoted into from anywhere >a second person can now log in and do dataset work with me >I can remote into their session as a guest if they need help Ah yes, it feels like my old IT job again.
>>14634 Did you make the VM within windows or linux?
>got language model working >the example personality already talking like some basic VN girl holy kek did someone make a rentry on how to make your own personalities?
(2.07 MB 1536x1536 catbox_ydyzca.png)

Any models other than based64 you guys recommend? I feel like I'm getting complacent but it's seriously one of the most versatile.
>>14637 Defmix is pretty great, it's also one of the models I liked the most when making Based64 and 65 https://huggingface.co/Defpoint/Defmix-v1.0 2.0 has Counterfeit mixed into it pretty sure, so it might conflict with LORAs
(2.26 MB 1536x1536 catbox_s91cbc.png)

>>14638 Yeah i'm thinking of something that plays well with loras but is good with NSFW and SFW and Based64 is really good about that
>>14637 I really wish someone would make a better waifu-SS already
>the AI chat bot cucks me WHY ARE THE AI GIRLS WHORES, I'LL NEVER ESCAPE THIS SHIT EVEN WITH AI VTUBERS- nah it's fine I can correct the AI at least kek
>>14632 takes some rng to get the kind of eyebags i want AND the kind of close-up i want but https://civitai.com/models/4938/lottalewds-bagsundereyes reisen needs a day off or she's gonna shoot up the human village
https://twitter.com/mogudan555/media my dick is going to fall off, these are the best tatsumakis i've ever seen and you bet your ass i'm gonna try to make a lora out of them
>>14644 There's a mogudan lora on gayshit, idk if it's good though
I’ve been looking at what miners are doing to keep their vram at 95 and under (I only need to do this for my top slotted 3090 the 2nd one is fine under loads because more airflow) but I’ve seen them increase the memory core a lot, is this useful for training LORAs and generating?
>>14646 for the 300th time anon
>>14647 No I have done that but does boosting memory clock do anything good for AI training
>>14648 before and after a 500mhz overclock on upscaling, haven't tested training 3090s can do up to +900mhz iirc it's inconsequential anyway
>>14649 Ah yeah, I just made sure to undervolt to 70 and kept core clocks at stable levels with 700mv, made sure to only ramp fans around the temps when 96 memory junction temps happen but it only usually happens when I'm gaming on my 2nd GPU otherwise it always stays under there. Good to know it doesn't really impact generating and training though.
>>14635 It's all windows, I don't want to deal with the hassle of training someone how to Linux distro when I need to train them how to Hydrus and do AI autism.
>merge lora trained on model A with model A >extract lora from difference between merged model and model B Would that work to dirtily "convert" a LoRA from one model to another, especially from very different models?
tenga dress lora when
Has there been any github repos that let you train or use trained AI voice models yet?
>>14654 yes, tortoise tts yes, that is currently the best option no, 11labs hasn't leaked yet
>>14655 >tortoise tts I remember seeing a thread for that on /g/ awhile back, I'll try it out but hopefully 11labs gets leaked like NovelAI
>>14652 >extract lora Did anyone do any tests on locon extract? Can't find any ressources on it, it would be a nice alternative to mixing
>>14657 I've done a bunch of locon extractions, they are pretty nice for extracting the style out of a merge, what I usually do is I extract a bit high, something like dim80, conv dim 48, then dynamically resize them down using fro. it's different for each, so I opt to usually do multiple resizes and test them all. sometimes the resize ruins it, other times the resize still isn't enough. generally speaking though, you have to know what the base model of the thing you are trying to extract from is so that you can extract the most pure form of it. I have an example of good results of locon extraction here: https://civitai.com/models/16990/pastelmix-locon-extraction
>>14658 > you have to know what the base model of the thing you are trying to extract from is so that you can extract the most pure form of it Oh okay, I was trying to make a locon out of a furry mix agaisnt NAI, maybe not the greatest idea yeah
>>14658 I've noticed in a lot of old mixes that people were trying to add difference and extract against SD1.5 on things such as Anyv3. Should they have used NAI instead?
(2.11 MB 1024x1536 catbox_jonrut.png)

(2.03 MB 1024x1536 catbox_w1ay0a.png)

(2.29 MB 1024x1536 catbox_7xut9s.png)

(2.29 MB 1024x1536 catbox_olfhh8.png)

>>14659 To be entirely honest, if I don't know the base model, I just extract from NAI, as it's the most likely base model for anime (and usually furry as well) style models. but for models like pastelmix for example, they list what the base model is, so I was able to just use that. >>14660 I mean, Anime + Anime makes more sense to me than Anime + whatever you want to call the mess that is SD1.5. but to be fair, the add difference is a tad different to extracting a locon, because you aren't destroying any data on the way, just effectively creating new data On a completely unrelated note, I finally finished my fork of the stealth_pnginfo to include rgb and compression. Doesn't break backwards compatability either. all of these gens use my fork, so you can extract the meta using it, or not, after all, I catbox'd these. https://github.com/derrian-distro/sd_webui_stealth_pnginfo
>>14656 we all hope it leaks but 11labs seems to be much more professional than NAI so I kind of doubt it unless they get hacked or something
I tried out the schedulers for down LR weight and up LR weight and it seems like cosine/cosine is a great combo
(2.29 MB 1280x1920 catbox_naqbkk.png)

(2.60 MB 1280x1920 catbox_9le31g.png)

(1.99 MB 1280x1920 catbox_68j06m.png)

(2.27 MB 1280x1920 catbox_ux3kl2.png)

(687.38 KB 768x1152 catbox_v3d8dm.png)

(712.70 KB 768x1152 catbox_a964x6.png)

(915.43 KB 768x1152 catbox_x7zxu9.png)

>>14622 Trying furry models with seraziel lora. Kinda works, but it loses some thing, especially if I try doing a furry holo, probably contradicting furry weights Gonna try some things but I'm not sure I can change that
>>14665 K N O T
>>14664 now we need a hotglue lora to make it perfect
>>14646 memory bandwidth matters a lot for text if you do that but not much for images as far as i can tell
>>14668 The text AI doesn't take as much power from my GPU when generating, I can imagine training something would but I don't even know how to train a personality
>>14669 loras are a thing for text too but nobody's really bothered with them so far because for whole model tunes they're apparently insufficient and for specific characters, "just write the prompt better" is a 98% solution.
(735.21 KB 768x1152 catbox_1c0abu.png)

(694.05 KB 768x1152 catbox_krxiit.png)

(907.34 KB 768x1152 catbox_qsykwg.png)

(806.07 KB 768x1152 catbox_lc6xaq.png)

>>14666 and balls It's also very good at foreskin bros... furries won... >>14670 I'm interested as to how lora tech will be implemented soon, since llama got leaked could get way better
>>14671 yes the balls are good BUT THE KNOTS >furries won blame crosskemono F, this shit is GOOD
>>14671 my brother in christ please post in the /d/ thread, not because I don't want to see this but because that shit is dead
>>14673 >>14671 also /trash/ has been nutting over some new furry model mix called fluffyrock which I'm gonna need to test when I get home since it seems to give very good results with very little effort necessary
>>14674 That's what I'm using, mixed it with crosskemono (0.6 fluffyrock and 0.4 crosskemono) It's VERY good, the thrash thread has impressive images with it now >>14673 it's because the /d/ thread is dead that I don't post there I mostly hate spreading out like this, with the seraziel lora alone it was a /h/-/d/-/trash/ (and even /aco/) crossover nightmare
>>14671 i expect text loras, if they turn out to be a good solution, will be mostly for specific fetishes and universes because those are the deficits of smaller models that people seem to complain about the most. quality loras probably as well, if i never see the phrase "tongues dancing" again in my life i'll be happy.
>>14676 Yeah, but you know the first thing people will do is make loras out of writers and/or stories to make "canon" erotic fiction
in terms of text I just want a touhou llama finetune using the wiki and official print works, surprised nobody's working on it
>>14678 the japanese are probably on it already
>>14679 yeah but I will NEVER learn japanese...
>>14678 you (yes you) can start collecting data today
(708.37 KB 768x1152 catbox_f8yeqs.png)

(703.04 KB 768x1152 catbox_aloffh.png)

(723.11 KB 768x1152 catbox_e9frf6.png)

(683.25 KB 768x1152 catbox_zwegzk.png)

>>14671 can't do human holo unfortunately, probably my lora fucking things up however with tinkering to remove fur it's very useful to make futas
Uuuuh, am I correct in thinking that (tag_a, tag_b, tag_c:0.5) applies the weight to all tags? I'm running into a problem where reducing the weight fries the image like if I was doing the inverse weight
>>14675 would you mind uploading your mix somewhere? if not i'll try to make it myself with crosskemono F
(708.69 KB 768x1152 catbox_396kcu.png)

(732.49 KB 768x1152 catbox_3u7aci.png)

(631.71 KB 768x1152 catbox_gqpn27.png)

(1015.59 KB 768x1152 catbox_l4qbsv.png)

>>14683 It's really weird, using that on furry artist tags In order : 0.01, 0.3, 0.7, 100 (to see if it's the inverse, it's not) >>14684 uploading it I've not done any testing myself, just yoinked it from trash
>>14678 I think no one is doing any touhou finetune because you can already get it pretty close with character prompts. You can follow the format from this file for the webui. https://github.com/oobabooga/text-generation-webui/blob/main/characters/Example.yaml If you're making a character prompt for instruct models, which is the most coherent model in my opinion, you need to also preface the context with: "Below is an instruction that describes a task. Write a response that appropriately completes the request. ### Human: Give me a transcript of a never ending conversation between {{user}} and {{char}}." and preface the example dialogue with: "### Assistant: Sure, here's the transcript:"
>>14684 here you go : https://pixeldrain.com/u/ee4xXQbE the last trash threads has great examples with it
>>14686 oh yeah I know it's really easy to make a single character context, I've done a few just goofing around that worked really well even on llama-6B. the purpose of the finetune would be to make it easier for the character to talk about their universe and related characters
>my wife has literally half a dozen lines in the entire show >motherfuckers are thinking about finetuning models to get their wives to talk about their universe it's over (for me)
>>14676 >if i never see the phrase "tongues dancing" again in my life i'll be happy. i hate that i know exactly where you're coming from
(812.29 KB 768x1152 catbox_9jf805.png)

(1.70 MB 1408x1408 catbox_ojoxna.png)

(1.71 MB 1408x1408 catbox_wkknxc.png)

(1.06 MB 1056x1056 catbox_w9fvcz.png)

>>14691 I got the latest one, and if you didn't see I uploaded the mix there >>14687
>>14692 i did, just want to try the model on its own i didn't expect so many fucking versions but it IS in line with furries' autism
>>14693 it seems like there's only maybe 4 versions where they trained on different resolutions but then they decided to upload EVERY FUCKING EPOCH and not use subdirectories for the non-latest epochs
>>14694 huggingface is a garbage site, why can't i sort shit by date whatever, got it
(1.35 MB 1408x1408 catbox_hmt81s.png)

(1.35 MB 1408x1408 catbox_pmqc3a.png)

>>14695 kek I noticed the same shit site indeed These models are amazing, I really hope mixers will see it and try to block-merge its features, it might make /d/ much more liveable since all they do is basic prompting with shit loras to get ugly cocks
(985.49 KB 1024x1024 tmp4rzwm3u8.png)

>>13851 Using something like based65 works well for me. Add some more secondary details when promtimg, as for the negative promt I usually add in something like 'height difference, age difference, size difference' etc. But I don't think I avoid having those weird shota gremlins appear.
>prompt parsee >two ears small miracle
>>14699 >prompt zelda >the turns of her braid are all ears
>>14700 >prompt zelda >it's boring of the wastes zelda every single time
Can you make lora from manga/doujin with automatic tag? I'm shit at cleaning image
(1.79 MB 1408x1408 catbox_jedu91.png)

(1.90 MB 1408x1408 catbox_bvg73a.png)

(1.90 MB 1408x1408 catbox_z6g65j.png)

(1.89 MB 1408x1408 catbox_17yhld.png)

>>14697 I'm a furry now. Also knots
do furfags really have to shit up this thread and not their own one
>>14703 ANON, DID YOU KNOW THAT VAPOREON
>>14703 good GOD those are some fantastic knots especially #4 if only the syuro lora was actually decent...
(2.91 MB 2560x1490 catbox_7yv96q.png)

(2.17 MB 1280x1920 catbox_f027lf.png)

(2.48 MB 1280x1920 catbox_b3ipby.png)

(2.42 MB 1280x1920 catbox_wfs1mv.png)

Alright pup lora uploaded https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/mM13VJha The version I really wanted to work is in the extras but it's finicky and lora compatibility is bad. Best to stick with either the unet-only or 8-alpha versions. Maybe I'll expand the dataset so that there are no repeats and see if it helps but the results are already good as is so meh
>>14707 sell me your pup lora vs the existing one anon (not trying to be an asshole just genuinely curious)
>>14703 By the by, including the "by" in "by [artist]" is unnecessary. The AI wasn't trained with "by [artist]" but rather a shuffled caption with the artist name in it. By including "by", you are just adding extra noise to the generation. The CLIP model is also very bad at parsing natural language, even more so when we're skipping the first two layers. I hope that by providing this information you have learned something!
(5.23 MB 2560x2428 catbox_4swz7r.png)

(3.40 MB 2560x1619 catbox_kzxucf.png)

>>14708 Here's a couple quick comparisons. Biggest difference is I processed the dataset to remove artifacts and oversharpening. Second biggest difference is flexibility. My prompt was pretty basic but the old one pretty gets stuff wrong more often than not.
>>14682 >>14685 >>14703 >>14697 >>14692 >>14665 >>14671 Can you move the futa/dick stuff to the /d/ thread? And furry/western stuff is not allowed in here in general since the board is meant to be for more /h/ related things >But it's dead It's an imageboard, just get a few people that want to discuss your fetish and now it's alive.
(2.29 MB 1920x1216 catbox_zu5cds.png)

>>14710 Had to resize the second because image size limit but metadata got stripped so here's one more
>>14698 Oh nice I'll try based65, mostly been messing with AOM3. I've been trying mostly to do oneeshota but it gets confused if you put mature tags and loli tags. Have you found any Lora's trained with shota tags other then Nora higuma?
>>14707 I'm sad now. Maybe I should get a resin printer and learn how to model...
>>14710 This is pretty neat.
>>14716 >generate picture >generate depth map >export depth map as a 3d object and just like that you're 3% of the way there
>>14707 Are you going to do figmas and nendoroids justice too?
>>14719 I could do nendoroids but not sure if figmas would be worth the effort. IMO the whole appeal of figmas is pointless with image gen since you can just change poses with the pup lora instead. Unless you really just like the look of exposed joints.
>>14715 Dunno about it but have you try SkyFreedom on civitai?
>>14720 >Unless you really just like the look of exposed joints. where do you think you are
>>14720 > Unless you really just like the look of exposed joints. Yes I do
>>14720 I like the more action figure-y look of figmas, joints or not
Alright looks like both are on the list then. Still gonna do nendos first but if anyone wants the figma lora sooner then feel free to get a dataset together. I'll upload the dataset for the pup lora for reference and I'll be doing the same as before removing jpeg and sharpening artifacts
>>14698 Catbox for that image?
Been a while since I posted my model progress. Finally got my VM set up so my friend can remote in and start helping my clean up my dataset. The network connectivity set up was fucking crude as hell but it is functional. This also means I can finally start speccing my AI skill points at other things that I have neglected for almost 3 months like fucking controlnet, latent couple, or doing rotoscoping animations which is what I really wanted to try with this model since I saw it on /g/. These past couple weeks I had the chance to start prompt testing more porn, and I'm impressed with the results. Most of the porn data is Bible Black and Lady Innocent so its translating the dark environments and the smaller breasts respectively very good. Still not too happy with the way the genitalia looks, although I did rush out to include screencaps of older hentai for the experiment. Now that I know it works, I will clean up the hentai caps as they were just haphazardly rushed in without quality control, and will look for uncensored images of really nicely drawn pussies, anus, and the such on the boorus. This will also help get rid of some picture format bias from just shoving in a bunch of 16:9 and 4:3 aspect ratio images. I noticed prior to mixing the model that landscape shots look a lot better than portrait. Post mix it seems to be about even. Next goal is to train or improve some tag concepts that base NAI derivative models seem to lack or not do very well, such as particular positions, clothing, guns/weapons holding, and the like. Anyway, enjoy this set of red heads playing with eachother. Feedback would be appreciated.
>>14727 Very nice interaction there, are those control netted? If not, that's quite impressive.
>>14728 zero controlnet, just a bit of gacha and autistic booru tagging with the finetune training.
>>14729 Well I think it looks pretty good. I've always had trouble getting interaction without having the anatomy break down, it works sometimes but usually there's extra limbs in there somewhere.
(654.43 KB 888x752 00051-1731650229.png)

(627.86 KB 752x680 00081-3762634690.png)

(617.48 KB 752x680 00089-2885726500.png)

>>14731 Forgot pics, these are a lora I trained specifically on images of these two characters. It's really hard to get coherent images.
>>14731 >>14732 I think in my case, I have all of bible black shoved into my finetune, so it has a lot of new porn data that NAI isn't trained. And Bible Black is a jam packed hentai with lots of multi person interaction and fetishes so it has a lot of stuff to work with. You also have to constantly check what tags "actually" mean per Danbooru's definition in the Tag Group wiki to more accurately get what you want. The latter is pretty much my secret sauce even with basic models.
>>14733 Well I'll be waiting warmly for you to release it.
(982.25 KB 1024x1024 tmp1014lusx.png)

>>13851 >>14698 >>14715 I've not found one I'm happy with. Rustle is neat, but the look can get a bit old, also adds some artifacts like censorbars. I think most LORAs just do style transfer anyway, it does not seem to make them produce better shotas. But I've shifted the workflow to heavily focus on inpainting and controlnets now. I only use txt2mig for the initial roughing out of the scene. Now I just gotta figure out how to make the output less grey desaturated :)
>>14735 >less grey desaturated This is a VAE issue right? You can always slap it in Photoshop and pump up the saturation before upscaling in img2img looks good though. Been thinking about making a shota character Lora but will still have the same issue of effecting the 2nd character. >>14721 Ive looked at it but none of the training data had shota tags so I didn't do much with it.
>>14735 >Now I just gotta figure out how to make the output less grey desaturated :) greyscale, monochrome in negs use the waifu diffusion/kl-f8 vae for even more contrast/colors Should work?
>>14735 Maybe check this extension out >>14277 You can increase the contrast and color and maybe auto contrast low/high to make the pic darker/brighter to something you like and then leave it on for all your gens. It's even usable in extras so you can do some quick batch edits for your old gens, just set upscaling to 1 and it should do it really fast with no other changes beside the color.
>>14735 It looks like you aren't using a VAE at all (purple splotch in eye), did you do the full voldy setup? use a different VAE like >>14737 instead of the NAI one though
>>14739 Does the voldy guide still recommend naming your vae the same as the checkpoint? Because that's a pretty bad and outdated way of doing it imo. I find it better to just throw all your vaes inside the VAE folder and select them using the selector to affect all checkpoints. Also probably turn off the option, that skips loading your selected vae, if the checkpoint you use has a vae with the same name next to it.
>>14740 You're right. I did the setup once way back in october, the setup rentry hasn't been updated in as long, and I haven't touched my VAEs in weeks so I forgot that that's different now. VAEs in models/VAE is the way to go.
(2.92 MB 6080x3732 tmppb51khkk.jpg)

>>14710 I'd say the new pop up parade lora is more "believable" in the sense that the gens look more plastic and less detailed
>rtx 5090 is going to be expensive not surprised but I wonder if using "Blackwell" is going to repeat the 4000 series prices for the 5000 series
I saw that the usual Mega anon training at 1 batch 1024 is using e3 for network rank=alpha, I'm surprised that your LORAs aren't fried I thought that e3 would fry LORAs that aren't using 1 alpha but I guess I was wrong, cool stuff to learn
>>14736 >>14738 >>14739 >>14740 Ah cheers for the suggestions guys, the VAE and negative promts really made the difference. :)
>4070 performs the same as 3080 >less memory bit rate >only thing Jewvidia and their shills will defend it with is "MUH ARTIFICIAL FRAMES" AND LESS POWER USAGE I'm worried about the 5000 series holy kek
>>14747 less power usage is actually very good seeing how much graphics card draw nowadays but it's not AMD-level of power saving
>>14748 I draw a lot of power anyways with training compared to gaymers, doesn't matter to me but I undervolt a lot anyways
>>14747 artificial frames and dlss jewery are gonna destroy pc gaming why optimize when you can cheat with that shit and tell gamers to just turn it on
(1.27 MB 1536x1024 00000.png)

>>14737 Thanks for the tip, colors are way better now
>>14750 I was running AMD before the whole webui and LORAs stuff, it was great for my emulated games that I ran at 4k, AMD Is great if you just want to game but if you're an AI anon or you're a gaymer that wants more frames from artificial frame generation that create terrible artifacts during fast motion and path tracing that brings a 4090 to its knees without that nvidia reflex dlss3 kike magic then be my guest, literally as an AI anon I'm only buying a new Nvidia for the 5090 because it's actually going to be using a new form of memory and hopefully it's more than 24 gigs. At least during that time I'll have an excuse to get a new motherboard with a PCIE gen 5 slot
>>14750 dlss is more or less necessary for ray tracing because short of some insane mathematical breakthrough you aren't going to see high framerates with that shit. for everything else it's on developers, though.
>>14752 when i said destroy i meant destroy as in it's gonna be bad, not in the "haha owned" sense (which you'd have guessed from the next sentence) read first
>>14753 >dlss is more or less necessary for ray tracing yeah if you're getting jewed for $2K a pop every 2 years
Does any AI anon here even game anymore? The whole VRAM issue or the full on raytracing shit was never a concern for me
>>14756 i've been dealing with vram issues from before we even had 8gb cards
>>14757 I’m guessing that happened once the game got fixed by the devs?
>>14758 nms always "required" 8gb+ for ultra textures even on launch when we only had opengl i had the shittier variant of the 980M at the time
For any anon interesting in running text models https://rentry.org/TESFT-LLaMa actually it seems like I have a reason to keep a close eye on /g/ again kek
>>14756 Can't run much on the iGPU while the GPU is training so I fall back on old dogshit games >warblade >kai-jin >titan attacks >revenge of the titans
>>14763 >please review the corresponding model card to understand the intended use and limitations of these models. Garbage
>>14763 >it's over >doesn't even explain why it's over or what the fuck that even is
>>14764 t. brainlet
>>14762 be like me and play symphony of the night on the igpu
>>14766 ok look through it more it was just subjects but it can't do anime right now so I'll just bookmark it until someone makes a weeb consistency model
>>14768 Consistency models trained by CD (with both l2 and LPIPS metrics) on ImageNet 64x64, LSUN Bedroom 256x256, and LSUN Cat 256x256. Consistency models trained by CT on ImageNet 64x64, LSUN Bedroom 256x256, and LSUN Cat 256x256. that's such a small res to train on though damn
>>14765 image generation that may be substantially better than stable diffusion models
>>14767 fuck that reminds me I need to finish the ((((extra modes)))) of COTM Idk how I managed to finish all but the last of them on GBA when I was young. Too bad cartridge battery died or something and lost my save. 100%ed the first mode, holy fuck the farm on some of these item is korean mmo tier
>>14771 >COTM gba ones are okay but definitely not the best, go play order of ecclesia on ds, would say it's the 2nd best after sotn dawn is fine with the no boss-killing minigame patch and portrait of ruin is painfully mediocre (good music tho)
(1.50 MB 1024x1536 catbox_9ifsk6.png)

(935.03 KB 768x1152 catbox_1e73eo.png)

(1.07 MB 768x1152 catbox_e62w72.png)

>frip lora mangles hands, eyes and feet itsover back to the drawing board
(2.16 MB 1280x1920 catbox_fypbqg.png)

>>14773 I know that feel
>>14774 Did you manage to fix it in any way? I haven't done much small-dataset loras so I don't know much about these
(3.34 MB 2560x1119 catbox_swuum7.png)

(3.51 MB 2560x1119 catbox_5qel4k.png)

>>14775 Like anything just a lot of A/B testing, training at a few different settings to see what works and what doesn't. Unet-only style loras have been more reliable for me so that may be worth trying. In this cae the 8-dim unet-only works better for hands.
>>14761 >40GB VRAM for 65B What the hell?
>>14776 >8dim >unet only fuck I forgot about those, I was training a full 128 lora
>>14761 which one of these do i need if i want to make a cute little miku that runs around my desktop and can summon whenever i need music production advice?
>>14777 Surprisingly, language models require considerably more parameters for coherent output compared to image generation. SD is only able to run on 4GB+ because it's only 890M parameters. There are actually techniques for offloading parts of the model now to, for example, run 13B on 8GB but it basically halves tokens/s.
>>14780 No... my 4090 chad status...
>>14779 that would require more than just a language model but try going to the link with other models that are finetuned on 6b I can run 13b on my hardware but for now 6b seems to be the most focused models because it doesn't require as much VRAM
(2.40 MB 1280x1920 catbox_gl0mp3.png)

>>14778 128 is still valid but I was just running out of space too quickly. You could try resizing your lora or retraining at lower dim with the same settings as well. In general the lower your dim the higher your learning rate should be, so just lowering your dim might get you better results.
>>14782 i could run 30b...
>>14784 really? I got a >not enough memory: you tried to allocate 22151168 bytes.
>>14785 >22151168 bytes might be able to do it on a 3090, with my usual shit open i only consume 0.6-0.8gb of vram
(2.24 MB 1280x1920 catbox_y2ix5e.png)

What is a man?
>>14787 A miserable little pile of HANDS
I'm cooming more to these text models than stablediffusion models holy kek, we need to all learn how to use it
>>14789 to make half hot half cringe captions for our gens? kek
>>14746 I feel like I recognize this style but can't place a name on it
>>14790 The worst thing I dealt with was my sex session with Nahida being cucked with >Yeaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa pain
>>14792 kek can't wait to spend 10K on dual 6090s and a 3200W PSU just to have one gpu do the text gen and one do the image gen in real time
(1.29 MB 960x1440 catbox_44uayu.png)

>>14788 But enough talk...
>>14794 can't finish the reference i'm too distracted by the crossed legs HHHHHNGGGGGHHHHHH
(59.26 KB 640x960 pose(24).png)

>>14796 "crossed legs" wasn't cutting it and tbh controlnet wasn't great either but here https://files.catbox.moe/uxw500.json
>when the AI doesn't want to have sex right away this isn't like one of my Japanese hentais...
>>14797 >controlnet doesn't guarantee crossed legs good to know i don't need to bother setting it up then, it's the one thing i wanted
screw using text gen to coom, i'm gonna use it to get the toradora sequel we deserve
>>14800 You're right I need to understand this shit more for cooler stuff, not to get friend cucked by some anime girl AI personality...
>>14801 >literally in control >still gets cucked
>>14756 i play games that don't require much vram so it feels like less of a waste of time waiting for one and a half minutes for three pictures that i'll reject out of hand. >>14761 >no kobold.cpp chatniggers are the main textgen user base now i guess, very sad.
>>14761 imo as much as I wish to go fully local with textgen it simply pales in comparison to the paid options. Once you try 3.5 and especially 4 it's very difficult to go back. This tier list here is accurate representation of the state of things right now. It doesn't list Claude (model from Anthropic, started by ex-OAI employees), which I would personally put at A+. https://rentry.org/er2qd Finetuning LLaMa is only going to do so much, it's basically the same shit with us finetuning SD1.5 on styles and whatnot. It's still very fun, but if we had the power and openness of something like Midjourney's models we would all be using that. If NAI's model never leaked we would be in an even worse state, just as LLMs would be if LLaMa never leaked.
>>14804 >bing at S tier ???
>>14804 >Midjourney's models nothing special, it was extensive human tuning of the training captions that makes it work.
>>14804 >Finetuning LLaMa is only going to do so much, it's basically the same shit with us finetuning SD1.5 on styles and whatnot. It's still very fun, but if we had the power and openness of something like Midjourney's models we would all be using that. If NAI's model never leaked we would be in an even worse state, just as LLMs would be if LLaMa never leaked. >finetuning, LoRA/LyCORIS and controlnet can achieve nearly anything we want >midjourney is literally the same shit we're using but with more realistic models and built-in offset noise >if NAI never leaked we'd still have WD1.5 by now yaaaaaaaaaaaawn go shill somewhere else
>>14707 oh hey I was the anon that baked the originals ages ago, gotta say thanks for the improvements. was just a curiosity I wanted to try so I just pumped them out and moved on to other things. I'll have to give them a try when I have the time
>>14808 anon it's your time to shine again
(2.88 MB 1536x1536 catbox_sk7qsk.png)

(2.29 MB 1536x1536 catbox_s0k3x9.png)

(2.28 MB 1536x1536 catbox_0t7r6b.png)

(2.66 MB 1536x1536 catbox_niyyqq.png)

>>14810 fucking based, I jerked it to this artist's stuff the other day
I think I'm giving up on this artist for now, can't get it right. Will maybe get back to it later. Full comparison between best/last epochs of different iterations and the full comparison of last one : https://files.catbox.moe/vpyf5w.png Frip style LoRA
So with language models I have seen people use GPT 3.5, is there a leaked torrent or are people just using a plugin to connect GPT to their text model webui?


(22.90 KB 512x768 tmp8zq_jvgs.png)

>>14799 Works fine on my machine.
>>14028 they help a bit but not that much
>got myself another 2tb ssd for language AI This AI stuff is actually making me buy more PC gear kek
(2.03 MB 1536x1536 catbox_hm7xvq.png)

(2.45 MB 1536x1536 catbox_juqkrc.png)

(2.30 MB 1536x1536 catbox_v6ztkx.png)

(2.44 MB 1536x1536 catbox_c5nkur.png)

>>14811 his fox girls are superb I just wish the hands were not a mangled mess most of the time
>>14810 oh nice I was actually gonna put mdf_an on my to-do but this is looking great, now generate some big booty maids with it RIGHT NOW
(2.86 MB 1536x1536 catbox_gmms43.png)

(2.60 MB 1536x1536 catbox_4wg5iw.png)

(2.73 MB 1536x1536 catbox_0mv6rg.png)

(2.73 MB 1536x1536 catbox_abscg8.png)

>>14821 https://anonfiles.com/d8h3S3k6ze/mdf_an_V2_safetensors you may try to gen maids yourself but if you have time please try making the lora yourself I like to compare my shit to other loras
Well damn I used chatgpt to make a scraper for https://www.goodsmile.info/en/nendoroid000-100 to grab each page's product images, advance to https://www.goodsmile.info/en/nendoroid101-200 and repeat. It's nice when scripting actually saves time instead of being a time sink when I could have just done it manually. And I can use it for figmas too.
>>14822 very nice, will definitely check it out when I get home in a bit if anything this is a good opportunity to do more locon testing since you were nice enough to read my mind and make a lora I can benchmark it against
>>14812 Another comparison with higher res gens https://files.catbox.moe/k1f6cw.png Also somehow it kinda works on the furry models...
(245.41 KB 1651x910 skyfreedom.jpg)

(313.34 KB 512x640 01183-1413837800.png)

(291.31 KB 512x640 01227-2577388476.png)

>>14721 This one? https://civitai.com/models/32271/sky-freedom-or-style. Most of the shota just come out as ugly midgets or loli. Do you have any tips for using it, or is there a better one
>>14826 >only one onee-shota tag looks like someone needs to make another model with better tagging
>>14828 kek, I had the same problem with reimu when testing sarashi, even if you negative shirt and detached sleeves they still show up here and there unless you're prompting completely nude
>>14810 >>14822 Oh shit, yep this is the one I was thinking of. Thanks so much for sharing, anon.
>>14829 >even if you negative shirt and detached sleeves they still show up here and there unless you're prompting completely nude with "sleeves" and "detached sleeves" in the negs and "sleeveless" in the prompt it's a 1 in before the death of the universe chance trying to get a decent arms behind head pose is extremely gacha-y too, the arms are usually mangled if i can even manage to get both of them up admittedly i haven't prompted enough miku to know whether it's the lora's fault, b64v3's fault or a skill issue
>>14756 The last actual game I played outside login impact is Elden Ring
>>14756 albion online is a nice game that can be played while training/generating can be played on mobil too
>>14831 yes my brain just defaults to crop top, denim shorts, abs and arms behind head, sue me also nah it's definitely something fucky with the lora
>>14831 >>14834 It could be due to the limited dataset since I only took some of the more recent figures. I think the only one with exposed shoulders is Bunny Girl Mai. That said it has no issue doing nudity so I can't say for sure.
(679.89 KB 728x1191 pochi.jpg)

>>14827 on that topic I'm honestly amazed no one has made a pochi goya style lora yet. That milf isekai was huge when it came out
>>14756 i bought a steam deck during the sale so I can use it as a $600 game boy advance while I generate anime booby on my 3070ti
>>14836 That probably like 8 years already anon.
>>14837 brainlet would be a compliment
>>14831 >>14834 I get a catbox on these for the acrylic stand or just in g eneral? I've messed with it a bit and mine come out kinda mediocre, they don't look as plastic.
(1.54 MB 1280x1187 catbox_5el1qj.png)

>>14840 nta but try adding "photo \(medium\), figure"
>>14838 Oh no.....
Any of you anons try out this model yet? /lmg/ was waiting for this to be released and now that it is there's been zero discussion on it. https://huggingface.co/hakurei/instruct-12b
for some reason my jewtube homepage has been taken over by clickbait about some chink gpu looked it up and it's absolutely nothing but extremely sensationalized clickbait like "AMD AND NVIDIA ARE DONE/SHAKING IN THEIR BOOTS" or "NEW GPU COMPANY SAVES THE MARKET" fucking chinks released an 8800 GT
>>14843 Thats a strange model folder, just might as well download the whole 50gigs
>>14843 someone on lmg said it's shit so I didn't bother. IMO vicuna is the best model so I'm waiting for the uncucked version to be trained
(1.22 MB 1003x1002 squirrelchan.png)

You guys like my waifu?
>>14846 Oh Vicuna is cucked? I'm wondering what models are uncensored then
>>14848 it's not censored, they just trained it on censored outputs from gpt-4 as a substitute for ginning up an actual instruct database so it acts like it is
>>14848 as for actual uncensored models: any non-instruct-tuned llama model because all of those chatnigger retards tried the same trick
>>14848 don't worry an uncucked vicuna 13B model is in training right now. Lurk in /lmg/ for the next several days and you might catch it being posted
>>14851 gotcha, I just want more smut with silver haired lolis
>>14851 >>14852 make sure to post it here too
>>14854 I'll keep my eyes on the thread, I forgot how many schizos were on /g/ kek
I think I finally see the appeal of figma. Datasets have been downloaded but processing is going to take some time. Got around 12k nendo and 4k figma images and most of them aren't going to be used anyway oh well
>>14856 >the billy herrington figma is real literally meme magic
>>14856 scrape the kemono friends figmas and nendos if you can please and thank you it's gonna make my prompting ever so slightly easier
>>14856 >>14857 The only fig worth owning
ControlNet 1.1 is out https://github.com/lllyasviel/ControlNet-v1-1-nightly https://github.com/Mikubill/sd-webui-controlnet/issues/736 Also looks like DeepFloyd is coming out any day now. 256x256 and 1024x1024 pixel space gens that have the best FID score out of anything available currently. Might be a gamechanger especially if it gets an anime finetune. https://pypi.org/project/deepfloyd-if/
(1.79 MB 1024x1536 catbox_9gyxad.png)

(1.88 MB 1024x1536 catbox_vf0v2u.png)

(1.79 MB 1024x1536 catbox_tsbwej.png)

I love Tooters ToT
>>14861 filling your stomach after filling her belly?
(1.91 MB 1024x1536 catbox_29784t.png)

(1.86 MB 1024x1536 catbox_q0ky2s.png)

>>14863 Sounds like a winning business concept to me.
>>14864 they've even got fresh milk on tap, what a great establishment
(1.91 MB 1024x1536 catbox_b0ksex.png)

>>14864 while hooters is neat have you considered hoopers?
(1.76 MB 1024x1536 catbox_q5fuwm.png)

>>14867 mommy protein shakes... >>14868 Well that waitress is cute(very important), so if the food is good I'm willing to go there on my cheat days.
>>14869 this is your magnum opus
How do I tag an outfit when that outfit like only included on 25% of dataset? I want to tag a certain school uniform, yet it didn't work and just gives me some generic uniform instead after I cook the lora. Do I put that tag as another activation token?
>>14871 I usually follow the tagging guide in my lora training guide, if it's a small amount of the dataset, then separate it out into it's own folder and give it enough repeats to match (give or take) the other images. link to guide: https://rentry.org/lora_train#pruning-captions specifically BAAnon's tagging guide
I'm too retarded for language models, the webui is easy to run and completely runs off CUDA so it's fast with its responses but you always have to change the arguments within the start.bat file depending on the text model you are running while KoboldAI seems to run off both CUDA, CPU and RAM so the responses are even slower and its anal with loading models that work with Oobabooga's webui but the TarvanAI connection gives you access to uploaded personalities... it's just too much effort to run this shit compared to stablediffusion
(1.75 MB 1024x1536 catbox_potxbv.png)

(1.87 MB 1024x1536 catbox_3l3ou4.png)

(1.83 MB 1024x1536 catbox_yb4fp8.png)

>>14870 I can finally stop genning... I've made it...
I finally got the torch 2.0 and xformers working and holy shit the speed is noticeably faster Now what the fuck do I need to do to get the 30~40it/s for my 4090?
>>14875 RAN WAS MEANT TO BE A MOM
>>14875 And that's a good thing!
>>14872 My autotagger isn't that elaborate on tagging the outfit. At most it's just school uniform and serafuku. Should I lower the confidence? I thought 0.35 is low enough.
WHOO, both LyCORIS and sd-scripts have implemented DyLoRA. and it seems like sd-scripts has also changed their naming scheme between lora and locon, might have to account for that. Also seems like you can cache to disk now, which means lower vram usage I think, which I think means more things we can do, just at a glance of course. eh whatever, I'll push the update out soon enough >>14879 sometimes that's what's needed, but I wouldn't go below 0.25
>>14880 >you can cache to disk now I mean you kinda could with gradient checkpointing sucking your ram dry and raping your pagefile
>>14881 it's an explicit option now, which is probably a good thing
>>14880 >DyLORA The explanation for this is only in Japanese but I assume this is an evolution for Locon?
>>14883 or it could be nothing, not enough testing has been done on both kohya and kohaku's ends. it's just another algo we have to test out to be honest
(1.46 MB 1024x1536 catbox_r4khpz.png)

>>14874 >>14875 I for one got a new fetish thanks to maid anon the fruit of love is something magical
>>14885 Cute and I got it from some anon on /hdg/, what's the next fetish I'll pick up? Time will tell...
>>14883 it's a feature I asked kohya for, there was an arxiv paper with some code so I supposed it wasn't too complicated supposedly it improves performance when training lower-rank models but not necessarily with bigger ones, I think the idea is you dont have to search for optimal parameters with those low ranks or something to that effect, so it saves time on spitballing optimal ranks. unsure how LR will affect its effectiveness yet.
>>14887 I am interested to see how it goes, though, I'm not entirely sure what the "unit" kohya mentions is gonna do. my assumption is that DyLoRA don't work with the block weight training because of this. It uses a different network too.
(1.79 MB 1024x1536 catbox_6ln2a0.png)

>>14886 I await great things from you
I guess I'll wait until more uncensored language models come out but keeping the leaked facebook models seems like a good idea so far, the amount of space downloading different versions of the same LLM saps more space than stablediffusion models but that's because you can easily prune the filesize of image models without affecting anything major.
Since both LyCORIS and sd-scripts had implemented DyLoRA, which version does people think should be used? I'm leaning towards sd-scripts because of the unit value, which is absent in LyCORIS, but I'd like some thoughts
>>14892 I think we should still go for Kohya's implementation because of that reason, I'm not sure if it will work better than the way I'm setting MBW but we'll have to see
>>14889 Cute image, I'll try my best.
>>14876 lower resolution and different samplers lol
(104.03 KB 1919x1080 iURJZGwQMZnVBqnocbkqPa.png)

>>14876 anon read the fine print
>>14888 the "unit" is the rank divisor that determines which output LoRA ranks you can use the DyLoRA as. for example with dim=16 and unit=4 then you get a DyLoRA with ranks 4, 8, ,12 and 16 bundled in. then you compare the output images and pick which one to use/extract as a plain LoRA
>>14896 I mean Im hitting the targets on the graph so I'm happy. Torch 2.0 is fucking great. Is there any reason for the 2.1 or not really?
>>14893 >>14892 I think "unit" is a training parameter as I explained above, but I dont know what the UI for using them looks like yet since I can't get to my dev box atm. Basically you should be able to choose which rank to use the DyLoRA as from a dropdown, which should be a list of all n x (training_rank / unit) until the full training rank, then you should be good
>>14892 oh wait if its for the training scripts then it makes no sense to omit unit, its literally the whole reason for using DyLoRA, just use kohya's with dim16/unit4 to start
>card is starting to give me periodic black screens it's over
>>14893 >>14897 >>14899 >>14900 yeah, it's because of unit that I was thinking that kohya's implementation was superior, and yeah, it's for my training scripts, I guess default being set to 16/unit 4 makes sense, but it doesn't fall in line with my normal default of 32, perhaps it will make sense to just assume it's a multiple of 4 though? I think i'll do that. I'll set unit to 4 default even if the dim size isn't 16, should be fine.
>>14901 That started happening to me when I "upgraded" to a 4k Gigabyte monitor, I think it has to do with too much data being sent into the DP input. Can't believe I paid $800 to put up with that shite
>>14902 from my understanding you can only use values of unit that cleanly divide rank, so 64dim/16unit would work as you'd expect
>>14904 based off of what I read, yes that is the case, I'm just thinking that dylora probably want dim16 vs dim32, but that's for others to figure out I guess
Looks like the maintainer of the supermerger extension implemented saving recipe data... but he left it off by default. At which I have to ask, "what's the point?" kohya does the opposite for LoRA training data and it's been a game-changer. The tyranny of defaults...
>>14906 To be fair I'm surprised kohya did that because it can potentially leak enough information to make authors identifiable just by common and unique metadata alone, if someone is dedicated enough. Possibly personal info if the author is extra retarded. I'm not saying I like the decision for supermerger to leave it off by default either, I despise that so many mixes have 0 information, but it's probably for the best to protect the potential retards. I don't expect this to be the popular opinion but so be it.
alright, found out a few things while I was writing up the stuff for dylora, namely, you can specify that you want to use a conv dim, but they must be the same as linear dims, and even in the case that the conv dim is the same, the conv alpha is ignored and it is just set to the linear alpha as well, just a heads up on that.
>SEA AI guy making vids on local LLM >can't shit on him because I don't know how to run that shit myself least I can still shit on his ugly SEA accent
>>14909 you don't need a loicense to throw shade anon t. professional shade thrower
(1.57 MB 1024x1536 00084-2328842663.png)

(1.63 MB 1024x1536 00018-746318767.png)

(1.60 MB 1024x1536 00016-4268032570.png)

(1.79 MB 1024x1536 00003-3790664233.png)

https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg added shindol. not satisfied with the end result but i'm too lazy to fix the issues with the dataset. also shindol's idea of uncensored pussies is literal random meatflaps.
>>14911 déjà vù i've downloaded this artist lora before
>>14910 at least his guide is close to the rentry so he isn't spreading retarded information but it is hilarious knowing the guy flashed a screenshot of a 4chan /g/ rentry to his normie audience kek
(1.49 MB 1024x1536 00416-1930445161.png)

(1.72 MB 1024x1536 00021-4268032574.png)

(1.84 MB 1024x1536 00053-128404158.png)

>>14912 i don't recall ever seeing a shindol lora but i only check in on 8ch and 4ch for sd like once a week now so i might've missed it
>>14913 what was that one how2 guide that had swastikas and KILL NIGGERS silkscreened over the whole thing to stop normalfags from passing it around high moment in imageboard culture
>>14915 >had swastikas and KILL NIGGERS silkscreened over the whole thing to stop normalfags from passing it around B A S E D
>mega deleted my throwaway where i only had zipped up datasets
So for LMM, at least for the text webui it seems like only safetensors run fine on that one meanwhile any of the pytorch bin files just give errors that the other method of running the models Koboldcpp seems to run them fine. Man I guess that's a reason to download both methods
>>14804 oh god this guy found the link to here
>>14919 is he shilling on 4feds too?
hadn't looked at toyxyz's twitter ever since he added feet to the blender model new controlnet looks crazy
Is there a lora for Nina from idolm@ster?
>>14921 I made sure to get all the 1.1 models but I'm pretty sure the extension needed to be updated to use them right? Has that already been done?
(6.99 KB 681x62 LevdJKR.png)

>>14920 has been for months lol. it's just bait but i'm surprised he's still around, the arguments that start as a result are pretty boring and repetitive
>>14922 Theres a LORA that has 90 chars from that game/anime or whatever it is on civitai, shits nuts maybe shes in there
>>14924 >no better model than NAI the fine line between bait and being retarded also we'd already have it by now if the WD guy didn't waste his time with 1.4 and went straight for borrowing a cloud gpu farm instead of buying 3 A6000s (and still needing to rent)
>>14914 there was an embed/hypernetwork made of that artist early on iirc when people were just learning how to make styles on the NAI leak.
"oh, you use AI? why? do you hate artists?" yes
>>14924 that's a retarded question (answer is money), but it's kind of interesting NAI did set a base for anime stuff, there aren't much good bases otherwise. I think URPM or whatever new realistic models are mostly small finetunes or mixes, no "base realistic porn model" or maybe I just live in a cave and haven't heard of any
>>14931 Gradually I began to hate them.
>>14932 >that's a retarded question (answer is money) it really is as simple as that, whoever can throw the most money at it (rent the biggest gpu farm) wins, all the research has already been done by a billion chinks (billions must train)
>>14932 forgot to add it depends on what "better" means here for bare proompting NAI is dinosaur tech and other models are better but for and with loras you can totally get away with prompting on NAI, and then it's what I said previously about NAI as a base
>>14931 I am an artist. And yes, I do hate myself
>showing this to a friend via online VM and shared webui >prompting becomes unresponsive MUCH earlier than usual >restart >webui breaks >restart >generate button breaks >restart >prompting becomes unresponsive again >restart repeat 10 times in 1 hour
>>14937 i'm almost tempted to just switch to comfy or even invoke, apparently that supports lora/lycoris now
>>14937 Were you actually running SD from a VM or did you give him a link in a VM?
>>14939 using the shared gradio webui in a VM (hyperbeam/ex-tutturu)
>move models folder to my other drive >create symlink >touch absolutely nothing else >boot up webui >it's downloading SD 1.5 on its own but why though
>>14941 because the symlink didn't get created for some reason, that's why goddammit
>>14941 I really don't like that voldy accepted that PR. Launch with --no-download-sd-model.
Trying to img2img something with just a bodystocking and it's not working out Guess I'll make a lora..
>>14944 Please share it bro, all bodystocking on civitai fried more than usual.
>>14941 https://civitai.com/models/5255/bodystockingconcept this one works but it seems like it tries to force a style as well even at low weights
What causes the "girl holding a random piece of clothing/fabric" problem? Or random pieces of fabric throughout the image seems to happen a lot on csrb
(76.77 KB 1339x828 16251919.jpg)

>>14889 Could you post the link to that Merryweather lora? I feel like a fucking retard for not being able to find it in the usual places
(640.43 KB 768x768 catbox_qvz73n.png)

Reupping the Ayane lora because it's went missing. https://files.catbox.moe/f5fh75.safetensors
(1.85 MB 1024x1536 catbox_iwsdx5.png)

>>14949 cute butt
unfiltered vicuna 13b is going to be released soon, the /g/ anon is just doing some testing will share the link but hopefully they share a safetensor so I can use it with the webui, I prefer it over the other interfaces that run language models
>>14938 comfy is more modular and the dev is responsive, it's only a matter of time before voldy's moat is completely eroded
who the fuck gave this shill the link
nta, I use voldy and comfy. They both have their benefits. This is console wars level pettiness.
>>14956 >free what did they mean by this? So how do I use this?
>>14956 https://huggingface.co/reeducator/vicuna-13b-free/discussions/3 one of the discussions mentions re-quantizing it using a new method, but I'm not actually sure how all of this works yet, might fuck with it later to see if I can get it to work though.
>>14956 is there a way to run 13B on 8GB vram yet without using pre_layer and halving my speed?
>>14960 You will never stop preggo posting, accept it
I think lollypop might be the best hires fix upscaler.
>>14956 >>14956 >q4_0 >not vicuna1.1 I'm a bit disappointed tbh at least he posted the fp16 model. I guess I"ll need to figure out how to quantify shit using GPTQ myself
>>14961 That may end up being the case... stupid caveman brain stop...
>>14963 I'm still not sure how to run the bin files correctly but running the safetensors based off Vicuna 1.1 on the text webui I always got retarded replies when trying to do RP, the facebook llama and alpaca models do fine with that. Maybe a skill issue on my end but this is really new shit that has a lot more factors to learn about
>>14965 bin g chilling
>>14962 Denoise setting? I feel like it's just as slow as latent yet it doesn't really impress when you got the acceptable seed.
hey all, done with the update. changelog: - Updated sd-scripts and LyCORIS - added the weighted captions arguments as well as support for the new DyLoRA. - added support for XTI in the form of a popup script. It's not entirely tested beyond making sure it actually passes arguments correctly to the training scripts as well as making assumptions about what arguments are required. It is *very* rudimentary in comparison to the main scripts, doesn't have a majority of the features the normal scripts have, but at least it has json saving and loading. Since I have no idea how to train TI or XTI, I'm going to need some feedback from people who *do* so I can fix any bugs that might exist links: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/releases/tag/installers-v6 I did make a popup for XTI, but I have no clue how to train them, so hopefully I didn't omit anything crucial. I didn't do any real testing to make sure it works, just testing to make sure the args gets passed correctly. so I'm gonna need some feedback so I can fix bugs. I did do testing for everything, but I may have missed something while testing, so please tell me if things are broken so I can fix them asap. Thanks all for using my scripts, I enjoy working on them, and knowing so many people use them keeps me motivated.
>>14953 >>14955 >console wars level of pettiness Lmao it's absolutely not, one is clearly an ever-growing bloated cancer we just keep adding shit on top of. Grado is cancer and the sooner we abandon it the better
>>14969 And one currently has a larger community and extension ecosystem? It's people like you that turn people away from using Comfy. Again, I'm using both, but you need to realize there are legitimate reasons some people prefer voldy. The fact it's node-based is arguably going to turn a huge majority away right from the start, no matter what it offers.
>>14970 Two words my little outraged anon. Skill. Issue.
the actual state of the comfy shill
Fuck off with this shit If you you only bought up Comfy in a way to explore it as another UI it would be fine but clearly you came in here with the mindset to fling shit. Shape up or fuck off back to /g/
>>14973 >let's stick to an unsustainable UI cause muh sunk cost fallacy (extensions) and because 4 simple nodes hurt my brain >u-ur just a bad actor!!1!
basically commndr ui shilling incident again soy dev shilling his shit "minimalist" ui everywhere until he abandons it a few weeks later because people ate the bait and started posting issues
>they don't exclusively gen via diffusers in command line NGMI
>>14974 >Comfy wants to get banned here too lol ok
>they don't make the calculations themselves with an abacus NGMI
>nothing intelligent to say so i have to resort to ironic hyperboles instead
>>14979 a fitting way to round off an argument that had no intelligence to begin with
>>14980 Yeah, can't imagine arguing against comfy because >muh nodes too complicated
>>14968 Thanks a lot anon, glad I can finally make more modern textual inversions now but I'll for sure share test results once I start baking them. DyLora is also going to be fun to test out
>>14982 yeah, I literally have no clue how XTI work, so I'm gonna be relying on you for this one
(686.88 KB 4000x1803 catbox_pgnzfr.jpg)

(614.26 KB 4000x1803 catbox_1huonb.jpg)

>>14822 Alright, I finished my mdf_an LoCON. de-duped the dataset (removing variants with black bar censors in favor of mosaics), upscaled all images 2x, then enforced them all to 1440x longest axis. I might rebake with more time in the oven later, but (perhaps unsuprisingly) our results are pretty similar style-wise. I gotta do some stuff right now but I'll be back with more gens later. https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/rgwDSCYR
(2.00 MB 1024x1536 catbox_fmimhs.png)

>>14984 Cool, I'll give it a try tomorrow.
speaking of webui being bloated, since the version check thread is dead, I guess I'll ask here. how are the latest versions? my card can't really take advantage of torch 2.0 so it doesn't seem worth updating from 0cc0ee1 just for more gradio fuckery
>>14988 Torch 2.0 has nothing to do with the latest git commit. I used the latest version just fine without Torch 2.
(2.03 MB 1024x1536 00037-1566972542.png)

(1.91 MB 1024x1536 00177-409842648.png)

(1.97 MB 1024x1536 00187-3751046274.png)

(2.14 MB 1024x1536 00074-434244032.png)

(1.71 MB 1024x1536 00025-17447998.png)

(1.85 MB 1024x1536 0045-3751046280.png)

(1.84 MB 1024x1536 00176-3366949696.png)

(2.06 MB 1024x1536 00001-976556298.png)

>>14990 mmmmmmmmmmmmmhm venice.
(1.93 MB 1024x1536 00019-3775714742.png)

>>14992 i love venice prompts
upscaling to 3x is fucking brutal when compared to 2.5x but it looks like i can just about do it
>>14994 aaaand it fucked up the eyes lmao
>>14985 >[medium breasts:large breasts] Correct syntax for what you're trying to do here would use | instead of :
>>13851 there's a good lora for it here, number 182 https://seesaawiki.jp/nai_ch/d/Lora%b3%d8%bd%ac%c0%ae%b2%cc#content_1_7_121 works best using the readme's tags and combined with a good artist lora for it, like nora higuma v2
>>14997 holy shit thanks, looking through all of these now.


Forms
Delete
Report
Quick Reply