>>30685 thanks for bake.

Fuck I wasted weeks over nothing because I'm an idiot. I thought dynamic prompts stopped working on my end, so I deleted Forge, switched to CompfyUI, and tried to get dynamic prompts working there. The versions for comfy are either overly complicated to plug or barebones, with no recursive wildcard prompting. After a long back and forth between having to rewrite my wildcards, reading up how hard it would be to write my own nodes, and retrying with Forge, I figured out my wildcards had faulty syntax, and I just had to change 2$$nextfile to {2$$__nextfile}. Shit. Back to writing my wildcards again.

>>30690 >Back to writing my wildcards again. i know this feel ;_;

>>30690 what's so good about wildcards? i tried them like over a year ago, out of curiosity, when i was trying every extension. are they for a quick fap session, mixing artists or what exactly?

I like it very much, although the videoization is still a rudimentary basic technology, or rather a poor one.

>>30692 nta, but having danbooru tags in a drop down menu categorized by tag groups is convenient.

>>30692 https://files.catbox.moe/g3q280.zip I don't use it in production but a varied wildcard collection is one fundamental piece of a good test suite You won't know otherwise if the new sampler / model / snake oil is actually good or good only at "one girl, {standing|sitting}, looking at viewer"

>>30684 My experience was they stopped having any visible impact at some point after v-pred v0.75, most notably furry negatives But it's difficult to tell these days as we have so many finetunes with undisclosed datasets

>>30692 Recursive wildcards are great for "image set" types of prompts. For me, it makes anything where you will have to implement variety into the mix at some point easier. Instead of having to think up tag combinations for various stages of stripping every other prompt or keep a list of prompts in a text file, you just write a master wildcard that takes care of that aspect and hit run. My PC is also extremely weak, so getting a great variety from just a single, specialized wildcard call makes it easier to leave the prompt running on its own. No need for me to babysit it and change the prompt stepwise every so often. >are they for a quick fap session Oh boy, am I jealous of you guys and your beefy PCs. I need wildcards because it takes hours to finish a few images, even without upscaling. I need to leave my PC alone while it gens, so I can't open up the fresh images while it's running or it'll take so much longer.

>>30699 Very nice

>>30700 thanks zankuro+fkey feels like a rather odd combo but I like how it turns out

>>30699 I have to admit, whatever they are doing at NAI is much better than the local regional prompter

>>30701 Yes but it's a very cute style.

>>30691 box please?

>>30704 https://files.catbox.moe/iadfra.jpeg

>>30699 >>30701 Really nice style. What does the actual prompt look like? High zankuro, low fkey, and oekaki?

>>30706 All those images have metadata

post more feet

>>30706 thanks but yeah the meta is all there just zankuro+fkey, nai4.5 seems to prefer sketch-looking style by default most of the time

>>30705 >schizo loras are you ok bro?

Why TF does Noob draw toes better than fingers? Do footfags have higher standards?

>>30711 >Do footfags have higher standards? uh, yeah? nobody has a hand fetish (no i don't care about the funny jojo man)

>>30711 Well fuck, I'm a footfag now.

>>30713 Welcome to the party. The good news is that a pretty sizeable chunk of the good and tasteful foot art out there is also loli, so you're already in good company.

>>30711 Toes move in much more predictable ways than fingers. They are easier.

>>30714 ehrm, what's the bad news?

>>30698 >Oh boy, am I jealous of you guys and your beefy PCs. I need wildcards because it takes hours to finish a few images, even without upscaling. I need to leave my PC alone while it gens, so I can't open up the fresh images while it's running or it'll take so much longer. Why not use cloud services then? Colab and Sagemaker last a few hours on free plan and it works at the equivalent of a 3060. Also I'm surprised that everyone likes doing image sets, I really like to do that too but the AI being schizophrenic with backgrounds really ruins it for me most of the time since changing a word or a line can completely change the background or composition of a picture which makes creating image sets really hard.

>>30716 The bad news is that there is no bad news which is bad if you enjoy complaining. I guess you could count most of FDY's NSFW stuff being paywalled, but it's on Kemono.

>>30713 just be careful, loli is an easy gateway for many more fetishes. i didn't used to be a footfag, armpitfag or femdomfag until lolis corrupted me. tbf everything is just way better on lolis.

>>30719 lolidom is so fucking hot I dont like femdom but there's something about being dominated by a small girl that makes it extremely hot

>>30720 lolidom is such a powerful fetish that it broke rimukoro after spending years restraining himself to writing fluffy SoL antics. there's something innately amazing about how you basically have to let it happen and your aggressor is adorable.

>>30717 >AI being schizophrenic with backgrounds This seems like something i2i or ControlNet could be used for. I'm already thinking of a few ways to do just that. Have you tried anything beyond frontal "text goes here" prompting? Like photoshopping a new pose over the old background and removing the outlines with i2i. Or making a pose with a 3D program and i2i that. Worst case, you could MS Paint a new pose over the old one if that's not too much trouble. Or prompt "{pose}, back background" (even with SD 1.5) and stitch that with a simple script onto your template background for a round of i2i.

>>30714 Thanks I guess, but I do like close-up images in general. >>30719 damn brats..

I genuinely wonder if the AI just sometimes ignores the tags or negatives to piss me off

>>30725 every time i bring up this issue people call me a schizo and just tell me "if you can reproduce older seeds then everything is working fine, it's just bad seed luck" bullshit

>>30721 Depends on the setting, I guess. BA's lolis are pretty strong from what little I know of the lore.

do you guys post your gens anywhere nowadays maybe there are discord servers where i can post nsfw (im okay with not being able to post loli)

soft or defined lolis? or both

>>30728 besides the usual places, not really. >>30730 soft because sometimes squishy.

>>30728 By the time the honeymoon phase passed, which has been well over 2 years now, I really only ever post publicly now for sharing technical information, not posting gens for leisure let alone clout. There's a lot less reason to do even that now though since local has stagnated. Some short bursts like Illust and Noob came and went. With NAI there's nothing interesting to share because it's using the same tool as everybody else. I'm sure a lot of my own opinion is shaped by being a dev though rather than consumer. I still cling on to threads I guess because of all the things I still maintain and I've met some genuinely nice anons due to it all. Also the (unrealistic) slight hope for something new to revitalize the local scene. But this is just a hobby of several and one that I still treat primarily as an outlet for masturbation, which likely due to being such an intimate act is ultimately why I rarely post gens publicly now. Far longer and sappy than I intended so now I will actually post some gens just to spare anyone reading my blog.

>>30732 shuuuuuuut uuuuup momoura

>>30732 are these 4.5?

Still hard for me.

>>30717 >Why not use cloud services then? nta but could you point me to where i'd find a good colab? also would there be a way to bring all my models/loras there without having to do it every time I start? That's kind of what turned me away from using it back in the sd 1.5 era

>>30735 that came out great. did you keep genning extra seconds off the last frame to get to 20 seconds total?

>>30735 that's pretty good

V4.5 full next week i hope it's not a trainwreck, 4.5 curated really didn't impress me

>>30732 The decline is unavoidable, I can't think of any reason someone without a niche fetish would still be involved in the local scene 2 years after the initial craze Most people are easily satisfied by vanilla porn of current season gacha / anime girls, thousands of artists can do that better for free than the best AI model For those who do have niche fetishes however generative AI must have been a godsend, personally it saved me a small fortune which otherwise I would have spent on commissions

I remember this one, people did something similar in i2i to gen dark scenes prior to v-pred models

>>30738 This method of creating a long video by saving reusable frames from a previously generated video as images and registering them as new key frame images. FramePack-eichi has a generation method that specifies key frame images. This can be done by combining this with LoRA.

Another beach episode.

>>30737 >nta but could you point me to where i'd find a good colab? https://tensor.art/ This site is good for casual prompting. Like if you don't fiddle much with the settings and just wanna gen random stuff then the site is fine. If you want to have an experience closer to local though, your options are colab and sagemaker. Colab is at https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb You'll have to download all your models/loras every time you start it and the equipment is equivalent to a 3060~ since it's a Tesla T4. You can get up to 6 hours of free genning a day on the free colab tier but the exact amount depends on how much free compute colab has that day. You can also make a new account and continue genning on that other acc for another extra 6 hours if you want to dance along the grey line and hope google doesn't mind but there's a non-zero chance that they might slap you for it. Fastest way to get banned is to do open two instances at the same time though, they HATE that shit so keep it to one instance at all time always. Then there's Sagemaker which is Amazon's version of collab. Upside is that I believe that they have Persistent storage, downside is that you gotta apply for it with a real phone number and an amazon account. The idea of having to tie my personal info to my degenerate gens always stopped me from doing so but sagemaker should be a superior experience than Colab if you can get it going. I think they offer 6 hours of free genning a day too but can't confirm.

>>30747 Thanks, don't suppose theres a way of installing extensions into the colab setup is there? The faster gen speeds are definitely a draw i'm interested in, but I used a lot good bit of extensions on my local setup too.

>>30748 You can install extensions but some might break. Also, if you aren't using a cloud service with persistent storage then you'll have to reinstall them all when you start a new colab.

>>30749 ah okay fair enough, thanks.

>>30743 neatorino.

>>30740 4.5 curated for me fixed all the issues I had with v4 that made it unusable. it's like what v4 should have been for me which is pretty cool. it also added a lot of characters from seasonals and negpip, so that's nice too

>>30741 AI is generally a godsend for more obscure non seasonal/gacha characters.

>>30752 4.5C doesn't have some of the artists that 4C has so we'll see

Where do you guys post most of the time? I always thought most of you were from half-chan’s /hdg/, but an anon said you were from /b/ degen, and now I'm not sure. I usually post on /hdg/, but lately I've been posting less and less with how shitty that thread is nowadays. Also having to catbox all my genss really sucks

>>30745 >tree shade Funny enough now I see a use case of this technique everywhere

I would post it here if loli was the subject matter, I posted a generation of a boy having intercourse with a woman in StraightShota(/ss/)'s ai thread and it was blocked, where should I post it?

>>30756 >2nd image what the fuck? don't tell me that it actually works

>>30728 >>30732 >>30741 >>30755 i've been self-uploading to aibooru since the nai leak. based on recent image scores, it seems there is a decline in interest from users. the novelty has certainly worn off, and there will likely be never a jump in quality we got when pony first came to replace 1.5 local video generation seems to be up-and-coming, but everything i've looked into can't do video gen on fp8, so i think i'm shit out of luck without getting a new gpu. there used to be various discords from big-shot model bakers on civitai, but discrod is cringe and most people posting there just use it as a dumping ground, was damn impossible to actually talk about the technology or hobby with anyone else. ironically, this board is probably the most social place when it comes to hentai diffusion

>>30754 Which artists aren't working? I know some like fizrot won't work because they have mainly nsfw pics, but I think they didn't work on 4.0c as well. But yeah I'm pretty happy with 4.5c too, it feels pretty competent and how 4.0 should have been if they didn't rush it. 16ch vae actually gives the details you'd expect from it without noising pictures like 4.0 did, artist mixing works very well.

>>30758 It's in metadata, controlnet tile at 0.25 weight, guidance ends at 80%, should be enough to create a light source and some contrast

>>30756 That's neat

>>30755 Usually here or /vtai/ if there's a gen people might like or a request. Most of the time it's just trying to gen of an idea out of curiosity to see if it will actually work and refining it or taking it to its extreme/degen to the point where I'd rather not bother to post anything

>>30755 >but an anon said you were from /b/ degen That's a shitpost. This board was made by hdg residents back in early 2023 mostly to avoid jannies. It's literally called hdg, why would it be related to /b/? But the "original" /hdg/ has been a cesspit for a while now yeah, not even really worth lurking. I post to pixiv and here if I even have something to post (lately not much)

The trick appears to be skipping the first and last 15% of steps

ol reliable

>>30755 on 'da 'cord of course!

Finally bothered enough to do a comparison between the 2 sampler setups I used Left is Euler a simple,CFG 8.0, APG on Right is Euler a CFG++ simple, CFG 1.35, APG off

>>30769 hmm, they're incredibly similar

>>30769 Rem is a cutie. >>30771 cute butt

it's out if anyone cares wish they released VT alongside it though

>traditional media and its related medium tags are still broken one fucking job

>>30775 try using ridiculous prompt strength like, uh, 5:: :: I haven't tried full yet, but using very high emphasis usually resolved the issues I had on curated without causing issues

artist mixing works but v3 mixes aren't plug and play also no inpainting model yet, it's just v4's for now

>>30747 nta but very helpful thanks anon >>30759 >without getting a new gpu thinking about this as well but 4 and 5 gen rtx don't feel like good value. maybe the chinks will give us dedicated ai cards.

>>30775 Can you show an example?

Typical of halfchan, I have absolutely no idea if 4.5 Full is good or not

>>30775 It works. Increase emphasis like >>30776 mentioned. You push even further to values like 10 and it still won't break down.

is "artist:" just placebo? i cant tell if there's a difference with or without

>>30782 no, meta tags are not placebo but sometimes for whatever reason the output might look better without them

>>30780 I'm enjoying it a lot >>30782 not placebo

>>30783 >>30784 i just noticed as you posted seems some artist tags work perfectly fine without it but some definitely need it, probably just best to always use it

>>30785 >probably just best to always use it yep

>>30784 I really think the VAE this time around nails the refined detailed look without being grainy or overly smooth tbqdesu

>>30787 NAI's internal VAE? I've only spent a few hours today on it but V4 is making me consider subbing again. Currently just burning through left over credits from the last time I was subbed (years?) ago. the enhance button is hit or miss, most of the time it kills nice artistic details like trembling lines

Haven't bothered tardwrangling gpt or whatever to add the new "features" to gelbooru but here's a modified (dan)booru tag copier script that automatically prepends NAI meta tags on export (again the meta tags only work on danny-b for now) https://raw.githubusercontent.com/shoebillanon/Danbooru-Tags-Sort-Exporter/refs/heads/main/Danbooru-Tags-Exporter.user.js

>>30788 >the enhance button is hit or miss It's a balancing act, strength at ~0.28 ±0.05 to keep the image mostly the same, noise (to add detail) at 0.01 to 0.1 depending on how adventurous you're feeling people say 0.1 noise is too much for low strength values like these but you should just ignore them

>ehhh i dont know it seems just like v4 >realize ive been prompting in v4 the entire time holy fuck im retarded

V4 is fun. >>30790 >strength at ~0.28 ±0.05 to keep the image mostly the same works really well, thanks. I was just using the magnitude slider before. >>30791 the model select is easy to overlook

>>30792 yeah nobody uses the magnitude slider (because nobody can be bothered to figure out how it actually works)

Did they use real images in their dataset? If you use detailed background it looks like its just putting a filter over a photo rather than the artstyle

>>30794 photo (medium) images should be in (excluding real people ofc)

random plug n play promptslop testing definitely better than v4 but, a little too much sharpening in my opinion, they just need to recreate smea tbh

>>30797 going to sound like a retard for this but have you actually tried just straight up prompting "blurry" at various weights?

>>30797 enhance to 1x, 0.2 strength, 0 noise

>>30798 I don't think its about blurriness i mean more the general linework and detailing, i might have to fuck with artstyles more though

Lowering guidance to around 3 seems to be good for softer images So does enhancing from low res, rather than genning directly in high res, though high res adds more details; a little overdone in backgrounds

>>30784 4.5 is p good

>>30802 I'm liking this new Shuten arc.

>>30771 >>30772 would burry my face so deep that excalibur need to pull me out.

>>30804 which one of those in particular?

>>30780 It's better than V4 and has their best prompt comprehension yet, but I still think V3 is aesthetically their best model. This one dips a bit into the 'fried' territory and doesn't really look as good as I would like. V3 imo has the best color/rendering out of any model and neither V4.5, V4, or illust/noob managed to get it right. And I stand by that SMEA should be a slider rather than a sampler, there should be a way to blend/dial-in the effect.

>>30807 >there should be a way to blend/dial-in the effect there is, img2img after genning

>>30807 incredible how you can judge an entire model on day 1, how do you do it?

4.5 actually convinced me to put out for a month sub, it's pretty nifty. Have to agree with >>30807 on aesthetics, though. I know even v3 had some seed style consistency issues, and they obviously cleaned it up from the trainwreck that was v4, but it can still be very schizo. Excited to do more with positioning but it really does seem to be a better regional prompter, the yuri pic here had a few failgens of the android seams bleeding onto Jill...

I'd love to try 4.5 but can't justify the price given the fact noob exists

>>30812 I'm not using the leftover anals I have, drop your grooming app burner if you want a token.

>>30813 Thanks for the offer but I'll pass for now and maybe pay up later on if it has any staying power

style mixing 2hard4me tbh weirdly i was kinda growing partial to the v4 mix i ended up, cant really find my way to that or v3 look, yet

>>30815 Use numbers instead of brackets, it's so much easier to control emphasis that way. Style mixing is definitely harder to get right but I feel like it's because the model actually learned the styles properly this time around, V3's styles always felt half-baked (as in it might've learned an artist's shading and coloring *to a degree* but NEVER fully and it rarely strayed from the default anatomy) in comparison.

>>30816 yeah im trying that too I think the main problem is some artists cause the image to have some completely blown out lighting Though, in testing it seems adding "warm colored" and "glowing" in negs is pretty much a requirement to avoid that

>>30817 I always had glowing+glowing eyes in the negs on V3 out of habit tbh

>>30817 also to add to this, "overexposure" is a good negative

>>30815 I got something pretty close to your V3 style, aside from using numeric weights it's a huge help to play with the (pitifully limited) style quality tags.

>>30820 Also, negging "muted color" (or putting it in pos at -1) may help with mixes where the color gets faded instead of blown out.

>>30820 looks like its leaning more into wamudraws

>>30822 It's not super dialed in and there seems to be a good amount of the v4 seed variation. Tweaking the numbers a bit could get closer, though it will never be 1:1.

Artstyle consistency seems fucked if you try to do more than 1girl, it ends up more like v4 where it just randomly rolls whatever it feels like

models never prompt exactly the same. old reliable prompts will have to be reworked. but 4.5 has a lot going for it.

>wow i like how this style is coming out but why wont it show me the cunny? >realize i forgot to put in nsfw into my prompt >add it in >quality drops dramatically, not even the same at all so they just fucking rushed the nsfw stuff entirely then

>>30826 I noticed someone say that nsfw gets added to negatives by default, I assumed it would get toggled if you put it in positives but there's still one token in negatives with all the boilerplate turned off. Could it be a bug?

>>30827 I wouldnt doubt it, we'll have to see if they even fix it or decide to start all over yet again

>>30827 the "empty" token in the negs was a thing in v4 too, not sure about v3 it's a single comma iirc

>>30829 Ah, I see (though that's still a bug, no?). Genitals are definitely fucked up though, both male and female. Like, I'm getting some horror pussies that look worse than Illustrious 0.1. Wouldn't be surprised if the NSFW is undertrained.

also why the fuck does NAI still not have a uniform noise schedule or any samplers made after 2023

>>30830 did you enable the furry dataset by accident?

>>30832 No, but I'm also wondering if the furry/anime datasets being tags (and not being able to turn off the current one, even with API) is causing jank.

>>30833 you just toggle between them with a button on the site, not sure if there's a new API command for it they really should've trained a separate furfag model but i guess that doing it this way still lets people use artist tags

i think i found a good range to work with, just need inpainting to get properly implemented not really like the old style but, i think ill have an easier time with that via vibe transfer if they add it

>>30834 >you just toggle between them with a button on the site, not sure if there's a new API command for it Yeah, I'm aware. API defaults to anime mode, I'm using metachs. >they really should've trained a separate furfag model but i guess that doing it this way still lets people use artist tags It actually doesn't so I have no idea what the fuck they're doing. Here's some examples: >bittenhard: works? 30 pieces on danbooru, 1k on e621 >kemokin mania -> el-k: neither work, 623 works on e621 (danbooru tag "kadim" doesn't work either) >hideki kaneda: 8 on dan, 250 on e621, same tag on both, doesn't work

>>30836 nice prompt. is that source# target# stuff something the model was trained to recognize?

>>30837 Allegedly yeah: https://docs.novelai.net/image/multiplecharacters.html#action-tags I'm not sure how much it helps since it will still sometimes confuse the subjects even when using named characters.

>>30838 thanks, will try this out more

>recognizes macci (possummachine) at 72 images >doesn't recognize zangyaku-san at 279 >think it might be because the artist tag is propping it up >still incredibly scuffed ??? 4.5 is so weird. I've found some characters (like more obscure Kemono Friends) work as low as 20 but a popular OC doesn't at almost 300? On another note, I've found that putting -1.5::film grain, dithering:: in your prompt for more painterly styles is a HUGE improvement. I also briefly wondered if they trained at higher than 1MP since I noticed 2char regional prompts turned out a lot better in native HD, but that doesn't seem to be the case since anything MORE than 2char completely falls apart and hallucinates extras.

>>30840 I did notice that artist OCs seem to often need their artist in the prompt. "1girl, zangyaku-san, gomennasai, solo" seems to work okay for me, but to be honest, I don't entirely understand what she's supposed to look like. She has so many different outfits on danbooru. That's probably why the model struggles too. Macci seems easier to understand, looking at her images.

>>30841 >"1girl, zangyaku-san, gomennasai, solo" seems to work okay for me This works for me better after realizing I'm retarded and turning my CFG from 4.5 to 5.5. Her design is just "girl with short brown hair and brown eyes" so I'm still not sure why it doesn't work well with other artstyles. The variety in costume shouldn't affect that...

>>30840 cute tenma

just because a character has its own tag it doesn't mean you shouldn't use supporting tags to describe it we learned this shit in the v1/leak era

I read something interesting on the analtan discord. If you have an artist mix and the artists apply to different characters separately, remove "artist:" and put "-3::artist collaboration::", that apparently helps.

>>30844 I had literally never given Anlatan money until two days ago, I'm used to Noob and its mixes where using seried name effectively functions as a style tag. Needless to say, I have a lot to catch up on.

very important research

this is definitely distracting me from other shit more than v4 ever did, nonshit inpaint when

>>30846 >where using seried name effectively functions as a style tag never really worked that way on nai unless you pair it with "official style" or "official art" or both. copyright/series tag isn't strictly required in most cases but it helps a lot with low tag count chars and especially if the char and artist happen to clash. eg on v3 some artists/mixes refused to output 2hus accurately (if at all) unless you had copyright:touhou somewhere in the prompt

>>30848 >nonshit inpaint when i'm waiting for vibe transfer tbh, even if v4 was scuffed to hell and back its VT worked 100x better than v3's ever did

>>30850 v4f threw it off with how scuffed it was imo, using it with v4c though is surprisingly good as long as they dont fuck it up for 4.5 but they dont have a good track record so far

>>30851 >v4f threw it off with how scuffed it was imo yeah i meant that it worked 100x better than v3's despite how bad v4 itself is

>>30852 Oh yeah fair, hopefully it wont take too long

>>30854 yeah no i REALLY hate that keyframed morphing look that all ai animations have

>>30855 i genuinely prefer the early days of using controlnet to make videos frame by frame (eg early toyxyz showcases) because it looks more like shitty rotoscoping than what we have right now (yes i'm talking to myself)

neg presets and negs in general seem to impact the quality way more than usual this time around even when compared to the quality tags toggle

cosplaying works really well

>>30859 trying to measure the effective of negatively weighted positive tags vs putting them in the negative prompt is also nightmarish. it's a pain in the ass to figure out which has a better effect

>>30861 generally i've found that doing negative emphasis was a much stronger effect than putting it in negs

>>30860 is there ANY hope for lolcal at this point?

>>30863 im going too far

so what's the verdict?

>>30865 way better than v4 but, somewhat underbaked

well, spent the evening doing some experiments with wan2.1 iv2 480p 14B on a poorfag's gpu... and it's "possible," but it takes like 15 fucking minutes to do just 2 seconds of video... if i try to push up to 5 seconds it'll go to up to over an hour. i'm going to spend tomorrow trying to figure out installing xformers from source, since wan2gp requires torch 2.6 but installing the latest xformers from pip forces to update to 2.7. if any other poorfag has figured out how to speed up i2v i'm open to ideas

>>30866 i think it blows local out of the water when you're trying to do something more intricate. but it has quirks like anatomy being ultra artist dependent

>>30860 your cute wife is a great cosplayer, anon

>>30867 I found it not really worthwhile to do vidgen on my 8GB GPU. I followed this https://rentry.org/wan21kjguide and used wan2.1-i2v-14b-480p-Q3_K_S >>30854 the first one looks great.

>>30848 >>30858 v4.5 offers the best interpretation of size difference I've ever seen How gacha are these gens?

>>30868 >but it has quirks like anatomy being ultra artist dependent isn't this the case on local when using loras? at least it was in the 1.x days. like i've said v3 had a much more defined "default" anatomy and i still think it's because it never really learned styles all the way (which is probably also why style mixing was easier coupled with the CLIP bleed)

>>30873 using loras actually stabilizes anatomy across artist mixes (without dramatically affecting those mixes if done right). it's why lora merged checkpoints are more popular than using the base models.

>>30874 i must've been using some overfit-to-shit loras back then or maybe it was just based64

>>30875 lora tech improved a shitton from based64 days to current sdxl noob meta. and based64 WAS a lora merged checkpoint to begin with...

>34 images hey that's pretty good

Does anyone know what all of the "dataset" tags are for V4.5? There seems to be at least: anime dataset furry dataset background dataset Either anime or furry is on at all times because of the "mode" toggle.

>>30878 There are only "fur dataset" and "background dataset", the anime dataset doesn't have a tag it seems, so use "background dataset" while in anime mode.

>>30872 Not very, only if you're trying to do really weird angles but, at lot of rng elements could be fixed pretty easily if it had proper inpainting >>30869 she's the cutest

>>30880 don't know how i mixed it up but, meant this one for the third image

4.5f is pretty good at being a "more of that please" model (copy all booru post tags > paste > generate)

>>30882 imagine with vibe transfer on top too

>>30879 Ah, sankyuu. Didn't know if there were more secret ones since the docs don't say anything about them at all. >>30884 Your wife has a very cute pussy. Excuse me while I spend another 3 hours genning various lolis with cleft_of_venus on display.

>>30885 thanks bro, im getting addicted again as well The best use of time honestly

>>30861 >>30862 They're effectively different things, visual vs. semantic. It helps if you understand what CFG is and how tokens are processed.

>>30887 My understanding of CFG (in relation to negative prompting) is basicallt that I was told it is "almost but not quite generating an image containing the content and subtracting it from your positive prompt". Mahouko's explanation there makes sense, and it makes sense in turn why I seem to get better results negative emphasising specific visual elements like grain and noise than negative prompting them.

I've had a lot of success with obscure general tags and characters in v4.5f, but one I cannot get to work under any circumstance at all is glands of montgomery. I wouldn't be bothered if Noob does it without issue. I went back to test with v4 and v3 and interestingly those didn't recognize the tag either. This tag is far more common (~700) compared to obscurer ones like presenting bra (~100) that do work just fine on v4.5f now. Maybe someone has a guess for what's going wrong...? 1st pic is Noob, 2nd is v4.5f. Simple prompt and Noob handles doing it with an artist who has never drawn them either.

>>30889 technically e621 does have more data (1000 hits) on them but they're tagged "montgomery glands". it could be that CLIP is pulling a miracle but I also feel V3-V4.5 have had generally undercooked NSFW knowledge that doesn't translate well between artists compared to Noob. Noob will render foreskin and proper vaginal anatomy far more often on the 1MP pass than NAI does, with NAI also being heavily artist dependent.

>>30889 >>30890 yeah its definitely underbaked, if you use the e621 tag it ATTEMPTS it but its kind of a horror show

>>30890 Huh. You clued me into that "montgomery" is likely fucking something up but it would be responsive to "glands". I just tried "spotted glands" for shits n giggles and it works perfectly. Maybe more is aliased we're unaware of.

>>30889 At first I was like "1.25? that's so low lol", then I turned it up to 5 and generated... Pic unrelated.

A bit gacha but honestly workable if i had inpainting

>>30895 Man, 4(.5) still LOVES adding greeble and scrimblos to backgrounds unless they're incredibly simple and/or landscapes. Agreed that I can't wait until we have the full tool suite, though.

>>30895 to be fair, im pretty sure its one of my artists adding glup shittos because it did that sometimes in v3, its just i usually can inpaint them out

>>30897 meant to reply to >>30896

>>30877 It's very obvious 4.5f has some additional datasets. it knows characters and concepts it has no business understanding. very esoteric ones at that.

>>30899 neo ranga?

>>30900 "neo ranga bodypaint" isn't even an actual tag but it works somehow

>>30901 CLIP + T5 does a much better job at grokking than CLIP alone. You can squeeze out some strange "pseudo-natlang" prompting by interpolating various tags.

>>30902 There's no indication that CLIP is involved, just T5.

>>30880 Good to know, maybe it's time to bite the bullet, hope it can do feral lemma better than base Noob The only thing stopping me thus far TBH is that I'm rather uncomfortable to associate my real identity to my gens through payment processors, fortunately I found some NAI account resellers on a local website

>>30904 >I'm rather uncomfortable to associate my real identity to my gens through payment processors if (((they)))'re after you that's not how they're gonna get you fam. they don't need an excuse either.

>>30905 nta but there are more ways your life can be ruined than feds. getting sloppy is the first step to trouble. however, you have to decide who to trust, and I'd say NAI is trustworthy based on their text gen clearly having some loli in the data set, at least when I last used it. they're also incentivized to not know what users are generating since the majority of the paying user base is some degree of degen.

>fearmongering *yawn*

>>30899 All of these have several dozen results on boorus. If you think the number is low on Dan then check Gel (included as of V4). They just did a much better job balancing training this time in such a way the obscure stuff is still recognized. Also if you think I'm bullshitting about Gel for whatever reason then try genning with "ijirare ~fukushuu saimin~". It picks up the hentai style. Compare that tag's results between Dan and Gel and you'll see why.

>>30908 That still doesn't explain characters like this one.

Can anyone explain what tagging fur dataset or background dataset does? I figure it'd be better to know for future reference instead of relying on intuition.

>>30909 Ah, well now you got me. That's actually fucked. I don't even have a guess as to where else they're sourcing from then.

>>30910 it... it does what it says on the tin anon... fur dataset (unnecessary, just click the petal button to enter furshit mode) enables the fur dataset and e621 tags, background dataset has pretty backgrounds + animals, probably at the cost of making worse humans

>>30912 okay thanks, i just wanted clarification in case there was anything to be wary of such as >probably at the cost of making worse humans

ass

>>30913 not even sure if that's the case tbh i've been meaning to test 4.5 more but every time i sit down i get distracted or draw a complete blank

>>30915 do cheesy porn scenarios with your favorite characters

>>30909 They must have run a more sophisticated version of this to generate dataset from seasonal anime https://github.com/cyber-meow/anime_screenshot_pipeline frame extraction -> character detection -> character classification -> manual naming

>>30913 so much this to be assful.

>>30918 catboxes?

>local is officially dead welp that's the final push I needed to kms, see you space cowboys...

>>30920 The death of local is a welcome one. Anlatan redeemed themselves after how pathetic V4 was.

go back to half cuck with your retarded console wars please

>>30919 https://files.catbox.moe/1jmumx.jpeg https://files.catbox.moe/nl93r9.jpeg

Beside the fact that there isn't any, do we need a replacement to catbox? Unlike half chan this site has much higher file size limit and also doesn't even strip metadata from webp, I no longer see the point

i haven't been here since the site was blue

>>30924 surely all the people freaking out about losing it will be able to help them out on their ko-fi lol

>>30926 just like how they donated to illustrious (not that you should ever give money to the gooks)

anyone know what would be a good way to get this sort of paintbrush texture? whenever I use fiz-rot it gives more of a 3d-looking vibe. I've tried "painterly" but seems to not do much in this case.

>>30928 i'm almost positive that's from his moemortem period back when he wanted to experiment more without putting in as much effort as the regular pieces might need to wait for VT for that

>>30929 aw damn, vibe transfer cant come soon enough then.

>>30930 we had constant problems with federating with baraag so i'm not sure how much stuff he put out as moemortem but had it been tagged as that instead of fizrot it might've worked to an extent

watching a 5060 review it's depressing

>>30932 Local isn't dying, it's being killed, the man who's killing it wears leather jackets

>>30933 yeah but i can't help but wonder if we'd be in the same situation had SD been invented a decade ago during the end of the "3/4gb of vram are high end" era

>>30932 i was planning on building a new pc when the 50 series came out but now im just eh why bother

>>30928 Like the other anon said it's a style variation, but I've found a few things help for prompting fizrot on 4.5. It seems to be more responsive to the newer tag "fizrotart", and you can neg Fizintine (and optionally Dagger) to help remove some biases.

>>30936 Oh and another thing that may or may not be obvious: remove dithering and halftone/screentone from your negs when promping a style based on fizrot. This applies to eg Onono Imoko as well and may get overlooked since these negs are now popular to combat unwanted texture. no_lineart helps too to get the "speedpaint" look.

>>30937 haven't seen a single soul talk about 4.5 negs or quality tags

>>30938 i feel like my gens are better without "very aesthetic"

>>30939 I think very aesthetic is mainly there to give people who don't prompt for styles something "fancy" to look at.

we're like what, 10 people here at best? we should get our own irc channel

>>30941 maybe as a backup but its comfy here

>>30938 Discussion is pretty sparse in general but I'm also not in the NAI Discord or anything. I pretty much get all my info from experimenting, here, and a few friends who have subs. Halfchan is useless for discussing NAI and I don't speak Korean to use arca. I do try to share anything I find interesting/potentially useful here.

>>30943 >NAI Discord for every even remotely useful piece of info there are at least 10k messages of nothing but pure slop, /b/-tier shit and normies who are genuinely too dumb even for the NAI web ui (the kind of people NLP is for) and the usual "it doesn't know x gacha whore from y gacha which failed 7 years ago and had maybe 10k players, unsubscribing right now!/give me my money back!” from people who don't understand how this works

>>30917 Right, this is intuitively what I would think was done as well in order to get that data in the first place, but if they really went out of their way to do that, I would have expected they'd advertise it. This isn't the same case as nsfw data, where that is questionable yet trivial to include. This is harmless to include but requires substantial effort. I would expect they only did it if it were trivial, which means the data already existed somewhere. Maybe I'm overthinking it though. Perhaps of note is that the autocomplete completes the tag as this, which no booru uses afaik: >luna (saikyou tank no meikyuu kouryaku ~tairyoku 9999 no rare skill-mochi tank yuusha party wo tsuihou sareru~) Also their autocomplete API shows there being 5,375 results for it. That is an extremely high number, and to my knowledge one they don't inflate. You can check the count for low-use general tags like "...!" and the count will roughly equal that of the one on Danbooru (580 vs. 548 as of now, slightly higher because of additional unique data from Gelbooru and e621).

>>30943 halfchan thread is completely useless for local as well, I got more useful advice here than in that shithole. there's too much shit to go through to find anything that isn't trolling or stupidity

>15 images >still not enough data to learn her sigh

>>30947 halfchan thread should have been nuked like 6 months ago when the noob discussion started dying down. it has been unironic months long flamewars and spam with maybe one person a day coming in unaware of the absolute state. when was the last time a thread lastest less than 1500 posts because posters would rather bait than just not post if they aren't genning?

>>30950 kek

GOD DAMN IT I JUST FOUND OUT very aesthetic IS A NEWBIE TRAP

>>30949 why would people keep shitposting instead of letting it go is something I don't understand yeah the progress slowed down, it's not as exciting as it was, but some people would sit there and shit up the thread for months despite clearly no longer enjoying genning. just let the thread be slow instead of turning it into this discord brainrot

>>30952 hi momoura

>>30954 hi

>>30952 I think it depends what you're going for aethetic seem to add more background elements if you include a background

>>30956 it felt like I was suddenly getting more coherent gens without it but now it just feels like I'm getting better gens either way. either I'm schizophrenic or the machine is bending to my will.

>>30957 welcome to NAI the seed generator is your best friend and your worst enemy depending on the time of day and moon phase

>>30957 Yeah its hit or miss, i think it depends on artists too just a quick slop gen but, left is without right is with

>>30959 left's anatomy is fucked, nevermind the background

>>30960 i said it was a quick slop gen tbh

>>30952 >>30957 I feel like their official example is the most relevant its ever been for this model in particular. Both look fine in their own right, but you harshly steer it towards what the majority of people perceive as aesthetic with those tags. If you gen anything outside of that then yeah, probably should seldomly include those tags. https://docs.novelai.net/image/qualitytags.html

>>30962 I did notice their comparisons of base vs quality tags looked very different from previous models. In a bad way. For some reason the model felt really annoying to wrangle without them at first compared to local but I have no issues now. I guess tweaking my negative was a huge help which is kind of weird that it's necessary when the model's been aesthetic tuned but whatever.

>>30963 >when the model's been aesthetic tuned Maybe quality tags is all the aesthetics tuning they do.

>>30964 I never really thought about it but that probably *is* the case, huh.

taking thing into 4.5c to inpaint seems to work well enough

>>30966 where's your wife's friend

>>30959 That pose is really fucking hot. If you have the time, prompt her getting fucked in that same position, please.

>>30967 here you go >>30968 I'll see if I can, haven't tried that in the splits pose yet

>>30968 A little gacha with the angles but, easier than I thought

getting ddos'd tonight, fun

>>30970 Lovely and very hot, thank you anon.

it's all about fun in the sun, right?

either i'm insane or 4.5 works much better at square resolutions than portrait/landscape just like v3 that or <1024 on either axis isn't enough

>>30975 >that or <1024 on either axis isn't enough hmm no, it still works best at square resolutions even above 1024 on both axes

>>30901 Thanks to you I know that neo ranga can output in NAI!

>>30977 Isn't there an anon that pops up in hdg and edg that does Neo Ranga stuff?

>>30867 poorfag's wan update. i got xformers installed, and it reduced gen time by 5 minutes (15 -> 10 for 2 seconds at 10 steps). i'm pretty pleased with the results. still looking for anyone else who is running WAN 14Bo on 8GB of VRAM to yield wizard knowledge

>>30974 very cute. especially the slime straddling.

T5 lets you get away with some pretty cool "tag interpolation" that pure CLIP models could never handle.

some anons mentioned that 4.5 feels underbaked with nsfw, but for me so far it's been performing really well. what sort of concepts did you have issues with?

>>30981 hnnng prompt is a bit gacha though

>>30982 i got skill issued at first but it's better now. you just have to fiddle with artist mixes and negatives, some artists play much better than others with nsfw. >>30983 >prompt is a bit gacha though picrel, for better or worse. I do think noob is much more consistent.

>>30981 Hot, I should check if my old nai account has any anals left.

>lowres, artistic error, (film grain:1.2), scan artifacts, aliasing, unfinished, worst quality, bad quality, jpeg artifacts, very displeasing, chromatic aberration, overexposure, dithering, halftone, screentone, multiple views, logo, too many watermarks, negative space, blank page, @_@, mismatched pupils, glowing eyes, bad anatomy, blurry, upscaled, bad hands, error, ai-generated, ai-assisted, dated, signature, censored, uneven eyes, artist collaboration, thoughts on this schizoneg?

>>30986 mine on the left, this one on the right, only using "masterpiece, best quality" for pos quality i think as long as you have a small handful of "actually important" negatives everything else is placebo

>>30981 >>30984 nvm, the prompt is amazing. shadow hands is something I didn't know I needed all this time. and the source# target# thing actually works

>>30988 KINO ALERT I've had angels and demons on the mind nonstop lately because of https://youtu.be/eTplxWaAD8o. Had it playing on loop for 2 hours last night, I think my brain is melting.

Any suggestions on how to gen Nahida without the hair accessory glued permanently to her head?

>>30991 pos: alternate hairstyle neg: leaf hair ornament not 100% but it does work

I'm still getting the hang of NovelAI, so I'm sure someone more experienced could do a better job.

>>30992 Thanks, that works.

>>30993 It does not let you get the clothes off?

man this website's a pain in the ass to post on right now hope things get straightened out

Not sure if there's a node out there that basically does this but in annoyance towards washed out gens due to certain artists/loras, I had chatgpt spit out a luminance s-curve comfy node to increase contrast (of a base gen so the img2img hires has a better starting point). If anyone's interested I can toss the python file in a catbox or something.

more of poorfag's wan experimentation... it seems if there is a lora the quality of the animation can be very high even at low steps... but if relying on the base model for complex animations quality drops pretty significantly

you will never guess what i've been playing for the past month

>>31000 skyrim?

>>31001 ah dammit, shouldn't have posted an argonian

welp back to escha i go

>>30995 The clothes are actually part of the prompt

>>30996 hot

>>30996 nice.

>>31007 i'm yoinking this mix. loli Lucoa is too cute!

my dick never recovered

>>31009 imp butt

>>31010 uooogh

refining

>>31012 great innovation in wife technologies

maid butt

>/here/ is basically dead for most of if not the entire day >halfchan /degen/ has actual realistic babycon >halfchan /trash/ is 90% mlp spam >halfchan /hgg/ is a bunch of training-obsessed schizos spiraling further and further into madness >halfchan /hdg/ is /hdg/ do i even bother checking the other ai threads here?