/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: 15.00 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 (Temporarily Dead).

Ghost Screen
Celebrating its fifth anniversary all September


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

(819.28 KB 1248x1824 1729790462273694-2545167176.jpg)

/hdg/ #25 RED Edition Anonymous 03/30/2025 (Sun) 15:43:43 No. 29491
Previous thread: >>28235 Resources: https://rentry.org/8chhdgOP
>>29491 unironically why didn't you make her pregnant?
>>29492 Not to nitpick but is rouwei linked just in case it becomes good?
>>29494 call it something between hopium and "well, it's *technically* an option"
>>29495 It's definitely plausible for the next version to be good I think. It just has to be less underbaked.
anyone have an idea what artist mix this guy's using? https://x.com/dakini_tencho I'm guessing a decent amount of wagashi
>>29491 Not sure it you want to include noob_v_29_checkpoint-e0_s4000 but I still have the original link though it isn't actively seeded The cyberfix and itercomp merges of noob vpred v1 may still be worth including https://huggingface.co/Doctor-Shotgun/NoobAI-XL-Merges/tree/main Also https://huggingface.co/Toc/toc/tree/main/models is the only place I can find something close to NoobAI-XL-Vpred-v1.0+v29b-v2 [88fc3afcef] >>29497 Looks more like mikoscrub
>>29498 >mikoscrub nice, I'll try that out
(251.20 KB 832x1216 1729790462273710-3530605094.jpg)

>>29498 I found 1.0+v29 here: https://huggingface.co/autismanon/sdxl_modeldump/blob/main/Vpred-v1.0%2Bv29b-v2.safetensors and added that along with the Dr. S itercomp/cyberfix HF. >>29497 Also, I'm fairly confident one of the the artists in this IS wagashi, given the textured eyelashes. This is the classic wagashi + healthyman with both at 1.0, for reference.
s~h~u~t~ u~p~ m~o~m~o~u~r-- ah, wrong tab!
kinda feeling the makings of an andy warhol meme
(707.56 KB 1792x2304 00024-354317580.jpg)

>>29503 Good gen premise.
(1.97 MB 3111x4000 grid-0001.jpg)

>>29508 It's consistent on my end, one of us might be doing something retarded.
This body type does things to me...
>>29508 >>29510 JOVI! :DDDDD
so this is the power of finding a sweet spot
>>29516 artists/catbox? that's some nice skin and face
>>29517 nta but I see some parororo and hotvr
>>29517 yeah sure https://files.catbox.moe/o2ckts.png it just tries to write one of the artist names sometimes with body writing tags https://files.catbox.moe/e7qp5y.png it's also more mature female biased, just trying to build some kind of dataset with it
>>29498 >the original link Can i have it please? my dumbass forgot to download the torrents after all this time
>>29519 thanks bwo
>>29525 Now do this
(4.42 MB 2880x1595 catbox_uowqyu.png)

if anyone ever tells you the base model is better because [incomprehensible gibberish about fringe styles] you are well within your rights to blow their head smooth off
>>29527 i mean, i don't think there is an objective winner here. all noob vpred models (not counting completely slopped ones) are pretty similar with their own quirks
(2.17 MB 1627x1632 02148-3771251836.png)

(1.33 MB 3997x2069 xyz_grid-0021-2295193457.jpg)

I'm gonna repost this here because I remember one anon asked me to share this lora but i just couldn't do it at that time because it was a shitty one, I hope he's still around here. It was trained on noob vpred 1.0 Also, I'm sharing this one. She's Monaca, an OC from the artist ShiShigaJ5. Here's his pixiv: https://www.pixiv.net/en/users/13302316. He used to post on /v/, and I think he still does from time to time. Her OC was partially responsible for turning me into a ped-- I mean, lolicon. I really, really love her, to the point that I downloaded koikatsu and her card just to take a bunch of screenshots to make this lora. Yes, it was trained only on koikatsu images, so it might have some issues. Besides that, I remember making a lora with this dataset about 8 months ago, but it was tagged really bad and it used to generate random stuff out of nowhere in the image. So I retagged it way better this time, with a confidence of 0.1, and spent a lot of time curating the tags. (v13 and v14 are loras trained with the old tags and v15 with the new tags). Without further ado here is the lora: https://files.catbox.moe/4xmkhi.safetensors and the trigger words: monacaN, short hair, [light brown hair:brown hair:0.4], blue eyes, short twintails, twin drills, hair ribbon, pink ribbon, thick eyebrows, Well that's all for today, thanks for reading my blogpost. Until next time, anons.
(17.39 KB dbooruUniforms.txt)

(1.36 MB 1152x896 00080-3767145828.png)

(1.37 MB 1152x896 00010-1507905106.png)


these wildcards were long so i'm sharing.
>>29527 >base model is better because [incomprehensible gibberish about fringe styles] i haven't seen anybody saying that base model is outright "better", just that it's more flexible which is correct. I use 102d and while it holds back some styles (painterly and sketchier ones), 95% of the time I don't care that much and stick to 102d now because it's much easier to use and I like its style.
>>29503 boxes?
>>29534 Catbox? That's some quality PVC
>>29536 Shit, why didn't I just ctrl+f the last thread, I knew it looked familiar. Thanks tho.
>>29543 very erotic
>>29526 Gave it a lazy try and I'm impressed my shitty prompt works, cheekjob isnt even a tag
>>29545 UOOOOOOOOOOOOOOHHHHHHHHHHHHHH
>>29542 very nice.
>>29529 thanks, loved the lora. Could you please share the artist mix or catbox?
(1.09 MB 832x1280 17040-20250221203453.png)

>>29529 Speaking of which, NTA but I made one based on the more realistic version of Monaca (from Honey Select 2) if anyone wants it I can drop a link as well. >>29529
>>29552 >if anyone wants it I can drop a link as well. Yes please.
(1.22 MB 832x1280 28579-20250406233105.png)

>>29553 This one is for Illust V0.1 though, (I'll train one on noob vpred 1.0 once I figure out how to rain on Vpred) the trigger tag is m0naca should be same hair and eye color, spoilered the preview image due to it not being very anime. (hope its okay) https://files.catbox.moe/i8h214.safetensors
Anyone who switched to lossy webp/avif, what quality settings are you using and have you ever encountered issues with inpainting/upscaling?
>>29540 nice
Has anyone ever made a quick rundown(tm) of pros/cons of comfy vs automatic-type ui? i never tried comfy but people always said it gives you "more control" but idk what that really means in practice
>>29562 Comfy pros/cons: + Hyper specific workflow creation and automation + Latest features + Developer friendly - Awful UI/UX compared to matured software (Blender, Houdini, etc.) - Unreadable raw metadata format (unless using something like https://pastebin.com/AJ6avY5f that) Invert for Auto and the other derivatives.
(1.67 MB 1536x2304 00010-626488031.png)

(2.36 MB 1536x2304 00021-4148060290.png)

(2.44 MB 1536x2304 00029-488477229.png)

(2.20 MB 1536x2304 00195-488477229.png)

>>29491 haven't been around since nai vpred 1.0 dropped. any new models worth using since then?
>>29566 no major developments unless you count naiv4. we've gotten a lot of mixes of noob vpred 1.0 that have made it much more pleasant to use. some of the popular ones are linked in the OP rentry but otherwise you can try dumpster diving civitai.
>>29563 I see, thanks. I think I'll just stick to reforge for now since I'm not generating too much lately
>>29569 imo there's no reason to use comfy if you aren't developing bleeding edge extensions or making highly automated workflows for some sort of slop mass-manufacturing
>>29567 i see thanks for the update. it's ashame that civitai is so full of trash that it's barely worth browsing but we all knew it would turn out like this
Innate artist mixing feels different on all illu/noob versions compared to v3 and pony. I don't get why. In my experience both v3 and pony seemed more capable of producing unique or interesting looking blends. Or maybe I just haven't used illu/noob enough.
>>29572 >I don't get why. I mean, they're all different models with different datasets, training strategies, and loss scaling. Artist mixing on Noob v-pred derivatives feels fine, it's just very sensitive and what artists are stronger than others can feel completely detached from their style or number of works. You can always try scheduled prompting, too: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing I haven't done it much myself but sometimes I'll tweak artists slightly for the upscale step or adetailer.
this is probably the only AI place I know that is not extremely cringe in some way, even if it's slow thanks for keeping it up anons
>>29574 i just wish it were a tad bit more active, i'd post more but I dont want to spam the thread either
>>29574 I'm amazed that this place is still around. >>29575 I'd post more if I had the time to gen as much as I used to.
(1.18 MB 832x1280 28417-20250405204401.png)

>>29574 Yeah weird how schizos took over almost every general on quad chan now, well here you know its so slow but comfy at least
>>29574 This place has almost no schizos, mostly because it's dead. The moment anyone talks shit about a model someone else likes, or posts a gen most people around here don't like, i.e., anything that's not 1girl, pov or kodomo only, they're most likely to start schizo fighting just like on half-chan. Don't think this place is superior to anywhere else. It's just another place.
>>29579 very cute
>>29578 >The moment anyone talks shit about a model someone else likes, or posts a gen most people around here don't like, i.e., anything that's not 1girl, pov or kodomo only, they're most likely to start schizo fighting just like on half-chan thread doesn't have to be a total hugbox. some arguments are fine unless they go overboard
>>29578 look, it's very easy: i've spent years talking to myself and turning my brain into a gay minefield full of stupid-ass rules that you simply need to not break or i'll shit myself in public
>>29567 >we're still stagnant fuck, back to my eternal slumber
>>29584 If anything major comes up I can implement it and have repo permissions to merge it on my own. Most stuff can be facilitated by extensions anyways though. You're already using hella old code by opting to use something like reForge since the Comfy backend is rarely synced with upstream at all, so I don't see merit to view this as a reason to abandon ship (yet).
>>29585 >If anything major comes up I can implement it and have repo permissions to merge it on my own. Thanks for the hard work. >You're already using hella old code by opting to use something like reForge since the Comfy backend is rarely synced with upstream at all, so I don't see merit to view this as a reason to abandon ship (yet). Fair, I just don't think comfy is comfy at all. Not even sure how you do stuff like inpainting with it.
>>29584 sad but yeah I don't see any reason to move from reforge for a while if you're only interested in SDXL-based models
I feel like I want to learn how to draw, I think. Any way to incorporate AI in learning?
>>29588 Gens can make for decent pose references and general design ideas. I'd be careful about using them to learn anatomy though.
someone hacked half-chan lmao
(223.17 KB 1024x1024 00057-2179003688.jpeg)

it was me
>quad chad hacked How in the fuck
who the fuck signs up to be a janny with their .edu address
I can't believe it's unironically over for 4chink. Anyway, my new home.
>>29595 it'll be back online soon enough will probably have significantly less jannies though kek
qrd?
>>29597 4chan was hacked, janny info leaked, some of the code was leaked, website is shut down by admins
>>29598 NTA but wow, they actually leaked source code, the absolute madlads
>>29579 >painterly, impressionism, [wamudraws], [opossummachine], fokkahim7, healthyman, tianliang duohe fangdongye, rustle, {{{asanagi}}}, [[[zankuro]]], lam(ramdayo), mika pikazo, aesthetic, digital illustration All this yet the output is 90% tianliang style NAIv4 style mixing doesn't feel right
I can't believe this happened
>>29600 Yeah its a mixed bag, you can take the same mix into a different prompt and itll change completely. The newer emphasis tagging works a lot better but, its still not the best.
>>29592 I'm so happy this happened. Fuck them, fuck the site and fuck their garbage moderation. The site deserves to burn in the fires of hell.
>>29596 >source code leaks >probably the captcha code too >it'll be back soon Don't hold your breath.
>>29604 they also don't likely don't have anyone to fix this mess seems like they were running a freebsd install that was EOL 9 years ago the soyfag had root access this will take a while to recover from
(2.23 MB 1248x1824 catbox_lc0plg.png)

(2.24 MB 1248x1824 catbox_tg8f8v.png)

>>29601 It's over...
God i'm glad the board is unlisted or I'd have to actually do work for once if /hdg/ flooded here
>>29607 /hdg/ is pretty much dead anyway, its only created and bumped by shitposters thinking they're trolling other people when in reality they're the only occupants; they wouldn't go out of their way to find alternatives otherwise they would have been here or other ai generals long ago on 4chan
>>29608 yeah /hdg/ had turned into pure crossposter shitposting cesspool ages ago, those early comfy NAI 1.5 leak days are long gone
Right now, I guess we are split up between here and treechan.
anyone getting some strange behavior with tag autocomplete on reforge? if i have any letter enclosed in parentheses with a specified weight, typing that letter elsewhere in the prompt yields no results for autocomplete. for instance, if i have (v:0.5) in my prompt, all of a sudden the listings for v (i.e. very long hair, violet hair, virtual youtuber) disappears from any autocomplete search results
>>29610 people still use treechan?
>>29612 Just today.
>>29610 there is another one tho. meta.4chan.gay
>>29614 if you're into 3dpd I guess
We don't need the cum splotches here, right?
>>29616 nah there's plenty /e/ type stuff here
>>29616 Don't post ilegal shit Don't be a dickhead Absolutely no realism or anything that could look like a real living person Other than those rules, go nuts.
good to see that you guys are still around no gen because either comfy or my gpu is shitting itself >>29608 yeah i haven't been to the other hdg in over a year far as i know its a ghost town entirely inhabited by bots and not much better >>29610 did a quick check and this place is vastly better
>>29605 Their 3-6 hour estimate for down time is optimistic. >>29610 There's an /h/ here too. It's mostly dead though.
>>29618 Maybe moving here isn't so bad.
(5.15 MB 2688x1536 comfyui_00012_.png)

>>29619 restarting my computer fixed it hands are still fucked but thats normal
rip 4chinks
(4.09 MB 1536x2304 00078-3974500329.png)

(3.07 MB 1536x2304 00112-1234826618.png)

(2.35 MB 1536x2304 00010-3407476158.png)

(3.43 MB 1536x2304 00052-1267493136.png)

nice to see this place a little more active with halfchan nuked >>29620 catboxes? nice style
analtan teased V4 VT with this, it looks way more flexible than V3 VT but you can't convince me that they're not gonna treat it as a bandaid fix for artist mixing
>>29625 It's released actually, they just haven't announced it for some reason.
>>29625 >>29626 Yeah, it's out. Now you can share "vibes" and other shit. https://docs.novelai.net/image/vibetransfer.html
i missed you guys
Ah right I forgor about this place, haven't been here since the pony days
For Civitai Lora previews are you "allowed" to use inpainting, adetailer, etc or just raw gen? Weird that you can't use their generator before publishing the lora if you baked locally.
>>29584 its over
>>29630 civitai generator sucks so much dude just use something like tensor art or run your own colab
>>29632 I'm just trying to post a lora for upvotes, but it needs sample images
>>29633 Can't you add images without gen data?
>>29625 trying it and it seems pointless
>>29626 >>29627 >supposedly major hyped feature is out >no announcement why are they like it i'll test it out later with some v3 mixes and other stuff
>>29636 like this* thanks autocorrect
How come people are still using nai here?
>>29638 Why not?
>>29638 I use it in tandem with local.
Haven't really kept up with anything since Noob's final release. Have there been any actual finetune improvements since, or is it just mixes? I vaguely recall some anon posting comparisons between noob and a further finetune he was doing but I'm not sure.
>>29641 some minor ones but thats it
>"modules will never return" >turns vibe into modules for image gen turk what the fuck are you doing
>>29634 I meant to ask if they should be reproducible
>>29644 reproducibility doesn't matter you don't even need to have gen data in there
>finally feel like prompting gura >she announces her graduation w-whoops...
>>29636 >i'll test it out later with some v3 mixes and other stuff Well, I like that it extracts more of the style this time around (but it's still not enough) and that setting the reference strength to 1 doesn't turn it into a glorified img2img. With enough wrangling, the new emphasis syntax and (multi-)vibe you could probably get in the same zip code as your old "complex" V3 mixes but calling this a bandaid solution would be delusional at best. I really hope V5 will be what they said V4 would be.
>>29648 Also, it sucks massive ass that you have to pay to encode images for VT even if it's just 2 whole anals but at least it might encourage people to share the encoded images.
>>29649 that's the main reason why I dont wanna bother trying to tard wrangle it to attempt the v3 style
>>29650 I get that but at the same time it's just two anals. *IF* gathering a bunch of single artist/mix pics, tossing them into VT and sharing the encoded files proves to be a reliable way to actually replicate styles/mixes we could have our own "have LoRAs at home" meme.
>>29651 Sure its two anals, per vibe image But, how many two anals are you gonna burn through trying to find the right vibe images+emphasis mix to get the look?
>>29641 All the idiocy surrounding it aside, I do think there's something in Illustrious 1.1. I'm playing around adding it to various previous mixes of older Noob or whatever. Dunno if it's a placebo or some shit, but it does feel like it adds more control over the prompt. Like making natural language more accessible or something.
Are you excited for Illu 2.0?
>>29656 if its better
>>28887 >https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg >Is the anon who owns this mega still here? I've seen some gens made with guweiz lora of yours but I don't think you ever posted the lora itself. Would be very nice if you shared it, thank you lmao is the guy who posted this still around? added it to the folder
>>29656 no, I think only latest vpred versions could really compete with noob checkpoints, but it's unreal to reach at this rate of their goal completion
>>29659 nta but how does illustrious 1.0 stack up to noobvpred 1.0 anyways? wasn't here for when it dropped but it doesn't seem like there's been much excitement over it
>>29660 it was pretty much just disappointment when it got dropped as on site gen exclusive on civ and others, because of their way to try make any money out of it, while the model is not really great at all, it's like, if there was no noob at all, yeah, it could be an upgrade over 0.1, but every other their release will be just inferior to a little fucked up noob vpred just because of the technology behind it and I personally think it's not worth it
>>29660 1.0 and 1.1 are only barely improvements
>>29656 Local has been on SDXL longer than 1.5 now. Just surpassed it last week. Pathetic how he's charging money for this outdated shit.
>>29656 People are actually still donating?
>>29664 maybe >>29664 >>29664 i'm suspecting that angels company is artificially raising the amount a little every day to make it seem like there's still support for the project. The individual funding goals for the models only combine to 1.2m stardust.
>>29656 >53 million memepoints for that """experimental""" shitbake lol
>>29652 Like three fiddy (you don't need to pay to change the emphasis, you need to pay to change the information extracted value and that's best left at the default for copying the style)
>>29656 reminder: illustrious is a kohaku fine-tune
>>29668 Remind me what a Kohaku is must've been a nothingburger if I remember artiwaifu, tofu, 4th tail etc but not this.
Didn't think i'd be back in this place (much less because 4cuck kill), whats up niggers, how has the last year and half been?
whats the new local meta? whats the new training meta? I want to get back into working on some unfinished projects Was working part time with an AI company but my gig finished and im back to neet maxing until something comes up
>>29672 illu finetunes/shitmixes same as before just with some new args and random shit to try. personally i've been liking lanzcos resizing for bucket since the sharpen effect helps the ai learn curves instead of blur, and network rank stabilization is just overall better for anything you're doing (tested both of these empirically to my satisfaction)
I took a deep dive on the illustriousXL models since finally some shitbag got enough good boy points to finetune Danbooru ontop of XL "correctly" (waifudiffusionXL lmao), but I saw that the base they used is a fucking shitmix on top of another as well. Wanted to see if I could probably tackle a clean finetune of XL with some tricks I learned from my gig.
>>29675 Yea, reading all of this shit pissed me the fuck off, its amazing the models are even an improvement to the peaks of 1.5
ooooh goooood, we have our very own "aibro" now huh. maybe the cucktainment site dying isn't a good thing after all.
>>29677 I used to post my stuff here before, I know how to behave.
(1.91 MB 1344x1920 20250416123830-1338840829.png)

>>29678 shut up momoura
who?
(1.14 MB 1024x1024 ComfyUI_35336_.png)

>>29618 >absolutely no realism Where do I put those? Does Civitai take them, long as they're gachaslut cosplays and not real celebs? I only know of Deviantart, but that place is awful both on a technical level and culture-wise. pic unrelated
>>29679 >おらなんご >龍分枚 what did she mean by this
Btw are there any news with Pony version whatever? I thought they were about to release it this spring?
>>29684 astralite and his cronies have released some preview pics but it's shaping up to be a bigger disaster than cascade. completely fucked hands, sepia rerun, the only thing it seems to do coherenly is ponies (lol). the fact v6 was "usable" was nothing short of a miracle.
(1.06 MB 1216x896 30320-20250414010742.png)

(1.28 MB 832x1280 29994-20250412015650.png)

(1.40 MB 832x1280 30002-20250412020310.png)

>Wake up >Not forced to read same old shitposting memes or dumb repetitive political ideology preaching yet again >life improves a bit Wow amazing how that works >>29682 Civitai will take them as long as theyre not IRL celebs or figures sucking dog penis or some shit, if you do wanna post shit like that, then im sure theres an /s/ copy here or another 4chan clone that might have a thread active I dunno.
>>29685 what exactly are these idiots doing wrong?
>>29687 >trainwreck arch they're only using to avoid paying for stability license >mixed dataset of anime/furry/rule34/pony/irl >captions are solely generated by an outdated ghetto rigged 7B VLM >project lead is a fucking moron, didn't understand the point of network dropout >"training cluster" is like 8 A100s in his garage >style clusters
>>29688 >mixed dataset oh no >autogenned captions with 7B model oh god >network dropout wait he didn't bother to set it? >style clusters huh? Overall yea this sounds fucked. And these are the Pony fuckers right? What about the deeper bullshit on the Illustrious fuckers?
>>29688 >network dropout NTA, but afaik it was not a thing with a finetune scripts? Only loras. Not trusting in that v7 shitshow anyway tho
>>29683 these gobbledegook signs are the only time i find myself grateful i dont know moon
>>29689 same thing he did before with hashes but with regular tags instead of random generated bullshit; the intention is to decouple words contained inside artist names from the tag for their style, which is kinda sorta a good idea on paper but practically it means you need to have a fucking spreadsheet open to do prompting instead of just typing the artist name for example, there's an artist who goes by "m k" who's impossible to prompt on any model without a lora because it's too hard for the model to lock on to just two letters like that
>>29692 i mean "v" works perfectly fine on most models. same with the expression tags like ":o"
>>29693 yeah but those have thousands of images as opposed to a few hundred anyway style clusters are fucking retarded
>>29692 Yea I can see how that shit would be absolutely hard to deal with on a full model unless you decide to just retag the name all together with one of his other handles as a fallback.
>>29692 ideally artist names would be inserted as unique tokens in the tokenizer like wdV (rip) planned on doing which would have sidestepped the issue entirely
>>29692 It's just to avoid mentioning artists by same same as in v6, legal and/or moral reasons. These prompt issues should be solved by the new TE, and if they're not then having hundreds of tags like style_cluster_187 would break things much worse.
>>29697 *by name
>>29694 yeah but it's certainly not impossible for a model to lock onto two letters as you said. the tag name just makes it more difficult for a model. pedantic, i know.
>>29697 sorry i had on my best intentions hat and you are almost certainly completely correct
>>29692 >impossible to prompt on any model without a lora There's this, but I guess it's cheating.
>>29689 >huh? I'm pretty sure he somehow clustered "similar" artists and all images from the same clusters get the same style cluster tag. From the 'cord samples, all the clusters look like (ai-generated:2) though.
>>29702 now thats fucking stupid
>>29701 thats vaguely accurate maybe it was just model issue idk man
(5.83 MB 1536x2688 comfyui_00001_.png)

sharkussy
So I guess this is the loli thread and the one on treechan is the regular
>>29706 this was the /hdg/ refuge from 2 years ago due to tranny jannies and some /g/ schizo, has nothing to do with with the 4cuck outage
>>29707 I will take that as a yes
>>29706 treechan has lolis too.
>>29706 you can post whatever you want on both threads as long as it is within the rules of the respective sites
>>29664 >>29710 nah, better to contain pedos here
looks like the shitposters remembered about this place
>>29713 >webps i will never understand why people use this accursed format
>>29715 >so where's the problem photoshop doesn't want to open them
>>29716 Man, we have to box here too? Treech doesn't prune metadata.
>>29718 yeah but all the schizos went to treechan
>>29717 weird, even shitty gimp opens webps
(2.22 MB 1536x1536 20250417091537-1338840827.png)

(2.08 MB 1536x1536 20250417091917-1990237440.png)

(1.96 MB 1536x1536 20250417091950-2362444119.png)

>>29722 there are too many threads now
(1.15 MB 1152x896 catbox_yclovb.png)

(1.15 MB 832x1216 catbox_567z85.png)

(1.23 MB 832x1216 catbox_zhae3g.png)

(1.11 MB 1216x832 catbox_0cfaiz.png)

New addition to my slop wildcard
>>29725 what, "toddler?"
>>29686 ><lora:wetbody concept:0.8> That's a Pony lora, is your model >ntrmix35NoobAI1.1mix2 a mix between noob and pony? https://civitai.com/models/600231/wet-body-concept-pony
>>29727 Nah, some Pony LORAs work on Illustrious fine its not all of them though I think it depends how they're trained, that model is just an earlier version of NTR mix I liked blocked merged over a later version of Noob (which I did because I liked the aesthetics but wanted to obtain the later Noob's knowledge) Theres no Pony whatsoever in that mix.
Why is it that when I use the catbox userscript I can see the metadata, but when I use an extension to view the metadata, nothing show up?
(2.81 MB 3344x11099 file.jpg)

(2.91 MB 3344x11103 file.jpg)

>>29728 It doesn't seem to work on my mix but thanks
>>29729 If by extension you mean "infinite image browsing", it doesn't support stealth png metadata The latest reforge has native support for that in PNG Info tab
(2.88 MB 8400x2100 Untitled-1.jpg)

>>29726 Now you guys are here maybe it's chance to do this again How low do you consider "acceptable"?
>>29732 not a fan of 5, 6 but not gonna sperg out if I see them posted *nta
>>29732 you know where the line is, do i need to post the chart again
>>29732 I'm fine with 5 tbqh and maybe even with 6. I blame kereno for it, his lolis (toddlers) are really cute, small and hot. But I don't usually gen them, I'm more into girls from 1-4
(2.28 MB 1920x1536 20250417130415-2362444123.png)

>>29732 feral lemma proportions don't weird me out too much because the legs are still long enough to mentally file them as gremlins rather than true toddlercon. it's when you dip into realistic proportions that it starts to get iffy.
(1.88 MB 1536x1920 20250417140127-3388332214.png)

(1.96 MB 1920x1536 20250417140314-1910886914.png)

>>29732 I'm more disturbed by how wrinkly and shiny the skin looks in 6 than the size. Other than that, all of them can get it.
>>29739 the skin looks absolutely freaky in all 6 images
>>29740 Yeah it looks glossy and shiny like the character got oiled up but I just dismissed it as AI things. It's when it gets all wrinkly in 6 where I get truly disturbed.
congrats, you all failed the normgroid test spectacularly! move back to 4cucks when it's back online, yeah? yeah.
>>29725 incredibly based
>>29742 geeeg
(2.20 MB 1920x1536 20250417142915-3462508044.png)

(1.92 MB 1920x1536 20250417143751-1984411012.png)

(2.13 MB 1920x1536 20250417145824-1797466464.png)

>>29745 Yeah this is some good shit, nice.
people training loras on wai is the new anylora-tier retardation
>>29747 jeets gonna jeet
(2.27 MB 1920x1536 20250417160825-1251797291.png)

body is no longer in use
(1.93 MB 1536x1920 20250417172744-263016832.png)

i tried boomerprompting for the first time and i genuinely feel dirty
why is this the only thread that doesnt auto refresh for me
>>29753 Are you using .moe? Switch to .se
>>29754 same in either case, just this thread which is weird
why are noob vpred models unable to differentiate "male on top" and "female on bottom" tags? it's almost impossible to gen small male on top of large female
>>29756 Didn't realize there's a difference between those two either. Also, do you mean some other models are better at it?
>>29745 Now I want big-tittied hag x fat, old batard for completion.
(1.89 MB 1536x1536 20250418052942-985067153.png)

(1.95 MB 1536x1536 20250418053353-919370962.png)

>>29722 Don't forget /slop/ on >>>/trash/
oopsies
>>29765 >Hugbox is the worst thing a loli enjoyer may come into contact >Hugbox makes people bold, bold people get caught Yeah ATF is the weirdest honeypot I lurk, the people who got caught there clearly got too apathetic.
(1.88 MB 1920x1536 20250418103346-964662853.png)

V4.5...
(2.78 MB 1536x1920 20250418112634-1338840805.png)

>>29769 >they actually fucking fixed traditional media and its related (medium) tags better fucking late than never, this is unironically one of my biggest gripes with v4
>>29771 now if only they'd fix artist mixing
>>29772 funny you mention that, i was watching the channel at the right time
>>29773 hmm interesting, we'll see if they actually do it
>8chan full of naiggers and pedos >treechan dead >4chan DEAD grim
>>29775 post a gen
>>29775 time to move hdg over to 7chan
>>29775 You can make an alt-/hdg/ on >>>/slop/. They have /edg/ with highlights and request anchor there!
>>29775 Hopefully 4cuck will be back up soon enough and you can go back to shitting up hdg again. Just spend those two weeks on sharty.
>>29775 This is not strictly a loli thread, anon. Post hags if you want
>>29769 Interesting. >>29771 I'd be more interested to see if they fix however they fucked the VAE. Prompt for something like "glitter" and compare V4 to local. It's so grainy compared to local.
>>29780 pretty much this, you can post whatever you want just don't be a dick and start bitching about pedo this and pedo that or you'll be the first guy to be banned in quite a while
told you that while extremely cathartic the death of the containment site is a bad thing in the long run
>>29783 >[extremely obvious thing] >told youuuu woaw
>le r*ddit snark
>single low effort baitpost >triggers hour long pedo meltdown How is this place different from 4chan /hdg/ exactly?
>heh, i am... le threadshitter :^) wow i've never seen this bit before
>>29786 4cucks brought 4cuck with them.
(3.80 MB 2688x1536 00180-3047995566.png)

Clowns gonna clown no matter which circus huh
How does one develop a superiority complex from using a dead altchan?
>>29789 Hopefully our janny here is better at cleaning up their messes.
Anyone want a piece?
/hdg/bros (REAL /hdg/, not the 8chan spinoff) is this our new home?
>>29778 >They have /edg/ with highlights and request anchor there AHAHAHAHAHAHA
(2.33 MB 1920x1536 20250418201302-2557528563.png)

>>29793 tail eyes and mouth are fucked on the ai one but the legs look better
(1.58 MB 1120x1440 00179-1069880236.png)

>>>51 kill thyselves
(2.58 MB 1536x1920 20250418203530-2339711248.png)

>>29800 Forgor to mention is still a WIP by some anon on /b/ >https://8chan.se/b/res/474864.html#474864
Any news since 4tard was killed?
where are my pillars...
(30.19 KB 500x500 monkeyspaw.jpg)

>i just wish it were a tad bit more active
(2.74 MB 1920x1493 00305-1965972420_cw.png)

Dead thread..... But everything else is pedo and furry shit. I hate it here.
Illu 2.0 is out if anyone cares I guess https://tensor.art/models/840365567832646761
>>29805 Finally, you gave us your approval, cathag anon. Welcome, even though you've probably been lurking or even posting here for a while. Now let's see how long it takes for brittle and oekakifag to post here
holy shit, i might as well start posting toddlers here to get rid of fucking tumor that is cathagfaggot
Dead thread..... But everything else is pedo and furry kino. I love it here.
(913.51 KB 1024x1024 ComfyUI_38725_.png)

>>29810 He's got a point though, somehow all the hagfags disappeared.
>>29811 hagfags were a psyop, altchans are not raided by feds yet
(2.21 MB 1536x1536 ComfyUI_38908_.png)

>>29812 Maybe just out for Easter, or chasing Civit buzz.
(2.32 MB 1536x1920 20250419063104-1535198282.png)

(2.21 MB 1536x1920 20250419063749-2339711249.png)

thoughts on r3mix vs 102d?
>>29815 i prefer 102d, r3mix has more ai shine. chromayume is what you want for better artist replication but it fucks up anatomy and details a bit more often (not current chromayume thats ass, the one from a few weeks ago is what you want)
>>29815 102d is too stiff and overfit. i like r3 more, and if your artist mix begins to feel too sloppy unironically use https://civitai.com/models/1452858/slopslayer-lora-shiny-horrible-average-ai-style-to-put-in-negatives at negative weights
>>29816 chromayume v1 feels almost indistinguishable from vpred 1.0
>>29818 okay but if you plot them against each other it looks somewhat better
(2.34 MB 1920x1536 20250419075918-2339711279.png)

>>29809 this but unironically scorched earth might be the only way to get rid of these tourists
(2.87 MB 1920x1536 20250419081649-3422915983.png)

Let's just get along...
>>29823 you can't get along with shitposters, anon how haven't you learned this yet after a year and a half of /hdg/ being an unusable cesspool?
>>29824 Just ban them when they actually shitpost. We're not at the mercy of a retarded /h/ janny here.
>>29825 pretty sure this board's mod only deletes spam and illegal/3dpd shit and said he doesn't involve himself in anything else
>>29826 I would have assumed rule 1 covers a bit more.
>>29824 I can't tell which one was the shitposter in that argument.
>>29828 do your reps
>>29828 if you try to invade established peaceful board and try to enforce your beliefs, i count that as shitposting
>>29822 based >>29820 fake
>slowly finding out the styles I perfected for /hdg/ appeal are not liked anywhere else it really was an echo chamber in the end
(3.56 MB 1664x2432 00083-4129363233.png)

>>29820 >>29822 very based, love me sum yuri
>>29833 >posting to appeal to a crowd of anons anonie...
>>29835 yeah...
>>29832 idk the doujin authors disagree
like seriously if you put madoka and homu together and add the spanking tag madoka's the top over 95% of the time
>>29833 >he tried to appeal to highlightschizo no one to blame but yourself
>>29831 Outside of a situation you made up in your head to feel the pleasure of being a victim, when has that happened in irl?
Madoka being the top should be pretty obvious.
I'm in the mood to test some new shitmixes, any recommendations? I have tried a few that has been mentioned in the threads already.
Not about imagegen but it's hilarious to see zoomio-normogroids mald harder and harder about AI in *every* field. I hope the creatures who were trying to poison the imagegen scene don't tell these ones that it's all pointless, the psychosis is too fucking funny.
>>29842 not really a shitmix but, have you tried illust 2.0? >>29807
>>29844 Not yet, but I guess I should.
>>29844 nta if i can't dl it it doesn't exist
>>29843 > I hope the creatures who were trying to poison the imagegen scene don't tell these ones that it's all pointless, the psychosis is too fucking funny. They're still coping and believing it works themselves.
>>29848 Thanks
>>29847 I can't wait until they start overlaying trannyfesto-tier shit over the image at 50-75% opacity
>>29848 strange it's not linked on civitai
>>29843 I'm not going to bother watching this shit but if they're talking about doing what I think they're talking about doing(which is "poisoning" it with super high inaudible frequency ranges) this shit purely does not work for music. You can just cut off certain frequency ranges from training and wow it's cured. Worst case whoever is poisoning it uses a mosquito tone or whatever which is technically audible(but won't be heard by "most" people) but even that can be clipped out and they'll just piss off zoomer and gen alpha consoooomers. This sort of thing does technically work against speech training where certain shortcuts are taken to dramatically drop compute requirement but for that shit the people doing the training are going to be using bespoke datasets so what does it even matter.
>>29851 It was posted literally few posted above >>29807 idk tho. It does something but feels way too wild. It feels like it's listening to a natural language better but is so unstable. I'll wait for some better shitmix
>>29855 that was tensor art and it didnt have a dl link
>>29852 I meant anti-ai stuff don't train etc etc but lmao that's fine too >>29854 >You can just cut off certain frequency ranges from training and wow it's cured. Not sure how they're doing it, I didn't watch the video because I judge clickbait exclusively by its cover but that's exactly what I was thinking, if that's how they do it then even the laziest negative IQ EQing job will completely nullify it. I already have to do that shit to fix terrible hi-hat/tambouring hissing all the time.
>>29857 tambourine* I can't type
>>29857 i only dabble in music prod but theres so much shitty mastering out there it makes me fuckin angy
>>29859 Can you hear this hell? It gives me a headache real fuckin' fast. https://files.catbox.moe/al3ewh.flac
>>29861 instant headache in the first 3 seconds thanks internet
>>29862 The real torture starts at 0:13
>>29843 ofc, it's the same retard who was claiming that there is no female producers because of anime boobs VST
>>29864 To be honest most "female" producers I've heard about in the last decade turned out to be trannies.
>>29861 tbh I listen to a lot of shit similar-ish to this all the time so it's not really all that bad. The retard was trying way, way too hard at the start, especially with the fader and they probably needed to dampen the bass when it comes in since the track focuses so hard on the highs but I've heard a lot worse. Not that it's good, just that I've heard a lot worse.
any good way to combat overfried loras? (same exact composition, bad prompt following) https://civitai.com/models/974110?modelVersionId=1615193
>>29867 remake the loras yourself
>>29868 uuh, no
>>29869 then suffer
After years of obsessing over a single post in one of /b/s degen threads, I think I finally figured out what model and artist Loras the genner used. Just need to tune the tags, and I can finally lewd my sleep paralysis demon. I'm so happy. I know it's slop, but it's my slop. The lewder tags are quite strong, the dress rarely even stays, I probably need to up the weight. When I moved some tags, the arcade tag turned into pixel art. Also, if someone knows good tags for ghosts, especially the nightmare fuel eyes on the second pic (or really anything to improve my slop), let me know.
>>29871 make a dataset for it and ill bake u a lora
>>29864 >no female producers because of anime boobs VST wat
>>29871 For eyes you could probably very easily train one of those "addif" Loras by training on a before/after pair of images.
>>29866 You can't hear it then, be grateful. It's something in the high 12-13Ks and it sounds like absolute hell.
>>29875 i do av for my church and i keep having to stop these people whove gotten their eardrums blasted out their whole life from jacking it up so high it clips
(2.43 MB 1536x1920 20250419124715-3737279455.png)

>>29872 That would be very generous of you. I gen on a potato so I have no idea what you would need for training. How many pics would you need and how detailed do the tags need to be?
>>29873 there was quite some outrage about nani by wxaudio and he made video where he was saying shit like "this is why there is no females in the music industry" or something like this
>>29878 100+ is optimal (don't think this should be too hard tbh), if you want to go the extra mile do closeup crops, same number as the full res images. for concepts, the hard part of tags is testing which ones already fit the best so you can take them over. i think you probably already have a good grasp on that so again this shouldnt be too hard. (i figure it's something like horror (theme), ghost, sharp teeth, black sclera, red pupils, something like that)
>>29879 I see, thanks. That's a cute plugin.
>>29879 >the phrase "giant anime boobs" is actually in the one sentence pitch you know, maybe this timeline isn't all that bad
(1.93 MB 1080x1576 ComfyUI_38955_.png)

>>29867 Reduce the lora weight for the base gen, down to almost nothing if you have to. Then bring it back to 1.0 for the upscale step. Since by then almost everything is locked in and you just change the style. Add a CN, anytest or xinsir union if you need to raise denoise higher.
>>29880 Sorry if you have to hold my hand for this, but it's my first dataset. >100+ is optimal (don't think this should be too hard tbh) I'll try, but it might take some time. Should I exclusively collect pics with the eyes I want, or should I throw in some pictures missing one or two components I'm aiming for (black sclera, that black haze, black tears)? I should exclude character and artist tags, right? Does the picture's size matter? Specific resolution?
(3.30 MB 2304x1792 ComfyUI_00020_.png)

so uh, this is the place now huh
>>29879 While I can't imagine being the kind of cuck who gets offended on women's behalf about shit like this, I can't help but have my doubts about the quality of this VST. I've never cared about what the VSTs I use look like, only that things are laid out in a logical and easy to use manner. Sometimes the fanciest, most eyecandy UIs are the hardest to use.
>>29884 the most important thing is to make sure everything in the dataset accurately represents the look you want, i would advise not to include things that are only partially what you want since it can poison your dataset. it will also help overfitting if you include different camera angles. if you want to tag them as well just make sure to check the images manually to make sure that all the tags you want are where they belong. i'd suggest just adding em after autotagging; if you want to go the extra mile, add image composition / camera angle tags. ideally you want minimum res to be the native training res of 1024x1024, i usually just trawl gelbooru absurdres images.
>>29886 imo it's actually very good, control is low to the ground but it sounds good without need of tinkering with it, it has r2r version so you can try it yourself
>>29887 >i'd suggest just adding em after autotagging I just took all the tags from Gelbooru. Bad? I trust the booru tags more than programs/AI at identifying what's there, unless it's some ambiguous tag.
>>29889 nta but remove stuff from the "Meta" section like cropped, highres, spoilers, etc.
>>29889 yes, it's bad, they are terribly inconsistent, and inconsistency is the enemy with ai. autotagging is more accurate if you want you can tag your way and i'll compare them, it's no trouble
>>29885 my pillar...
>>29890 >remove stuff from the "Meta" section Would have removed anything except the general tags and highres, I'll toss that as well. >>29891 >autotagging is more accurate At least then I don't need a booru for the images. I already went through most of the relevant tags on the boorus to look for such pics, and I remember I found maybe a dozen that kinda had the look I was going for. Maybe I'll try Pixiv or any of the reverse image searchers.
>>29891 >yes, it's bad, they are terribly inconsistent, and inconsistency is the enemy with ai. this. when it comes to tagging there's no "it's better than nothing", wrong/inconsistent tags are more harmful than not having tags.
>>29894 And that is why autotaggers are bad. They give you more wrong/inconsistent tags. Human tagged images may be missing tags due to laziness, but they usually won't be blatantly wrong.
>>29658 Yeah I am, lol. Thanks, the nun is hot Spoiler cuz kinda aco
>>29896 the rules don't say anything about aco or /d/ stuff, or yuri, etc. just photorealism afaik
>>29897 Spoilering aco is considerate, I appreciate it.
>>29897 >Remember, furry and western art styles are not allowed here. 'Western' refers to cartoony art styles, realistic art styles or art styles used or often seen in the western part of the world and not where the artist or franchise is originally from.
I was lied to apparently >>29618 Still, where are you quoting these from? The rules only say: >3 : No furry or western cartoons
>>29895 >>29895 >They give you more wrong/inconsistent tags If autotagger refers to the algorithm behind the Danbooru tag button in Forge, then agreed. I'm always getting a lot of false positives. >they usually won't be blatantly wrong Depends on the tag. When I made my wildcard list, I found some tags that are ambiguous without the tag wikis, and so people mislabel pictures. For example, the "breasts_apart" wiki explicitly says it's not for small chests in the wiki, but there's a lot where they are incorrectly tagged that way.
>>29903 does forge still use deepbooru? if so then that's old as hell and shouldn't be used as an autotagger these days
>>29903 There are various detection models, trained similarly to image generation ones. You should be able to set a confidence threshold so that it leaves things it is not sure about, instead if adding things that aren't there. Obviously the ideal way is to use both booru tags and autotagger, then remove duplicates and recheck every tag manually. It's not too much work for a dataset of 30 or so images.
>>29903 For stuff that's ambiguous or is usually tagged wrong, you want to go by the model's existing knowledge not what the tag wiki says. Unless you're specifically trying to retrain that one concept.
>>29903 deepbooru has always been fucking terrible
>>29904 >does forge still use deepbooru? Yes, as of the last time I updated it. >if so then that's old as hell and shouldn't be used as an autotagger these days What are the reccommended ones now?
>>29903 >For example, the "breasts_apart" wiki explicitly says it's not for small chests in the wiki, but there's a lot where they are incorrectly tagged that way. The problem is that auto taggers are trained on this very data, so they will make these mistakes as well, on top of the other easily avoidable mistakes that humans wouldn't make.
>>29909 that's why you have to go over tags manually after autotagging
>>29908 >What are the reccommended ones now? The best is wd14 eva02, but even that gives me very obviously wrong tags at 80% confidence every now and then.
>>29911 Might as well use the original booru tags as your basis if you're gonna fix them manually anyways.
boorudatasettagmanager is still good for actually messing with tags, the russians or whatever that make it have updated it a lot since a couple years
>>29909 >>29911 what matters is not the tag wiki description or the autotagger, but whether the model you're training on understands breasts_apart as also applying to small breasts
>>29913 again those are still worse
>>29916 They are better because they are less wrong. The main issue with human tags is that they may be missing tags, but auto tagger tags will also be missing a lot of tags, because the vocabulary of auto taggers is comparatively tiny.
>>29875 no, I can hear it just fine. I'm just used to listening to shit with a lot of stuff going on and listened to the track for more than a few seconds. Like I said, it's not good. Whoever is composing it is trying way too fucking hard and it sounds awful, the opening synth is garbage and composers should never fucking touch the fader ever but after that it's just really bright sounding eurobeat.
>>29917 danbooru users mistag shit all the time.
>>29919 >Whoever is composing it is trying way too fucking hard and it sounds awful, the opening synth is garbage and composers should never fucking touch the fader ever but after that it's just really bright sounding eurobeat. >one of the most iconic pieces of eurobeat with some very unfortunate EQing >"heh this is garbage actually" -bedroom producer commenting on the composition and not the actual issue Alright dude, glad you're not an actual producer.
I'm glad we have so many music experts here, goes well with my still images.
>>29921 eurobeat is trash so it would have been destined to be trash regardless. and if you think shifting left to right to left to right speaker at hyperspeed is ever a good idea you should probably kill yourself.
>>29922 nothing to do with expertise just how good ears you were born with
>>29923 >eurobeat is trash so it would have been destined to be trash regardless. >and if you think shifting left to right to left to right speaker at hyperspeed is ever a good idea you should probably kill yourself. Tell me you're a nigger, a zoomer or possibly both without telling me you're a nigger, zoomer or possibly both.
nogen general
>>29925 how does it feel knowing eurobeat has had less cultural impact than dubstep(or by extension, trap)? Which is even more funny when you consider that eurobeat is the trash pop genre of electronic music. It's not deep or complex - it's just discount rave.
>>29927 >how does it feel knowing that zoomio-normogroids don't care about the thing you love? Feels genuinely great, thank God.
>>29854 by the way, nightshade "attacks" both diffusion and classifier models (i.e. wd tagger) it accomplishes that by taking the image, taking a certain noise timestep, adding then removing the noise with "dirty label" prompt and then mixing this shit back into the original image it's as effective as shit ass tagging, especially if you choose "dirty label" to be completely random, if you choose to bleed infrequent concepts into frequent but similar ones, you'll have a much stronger effect. basically it relies on so-called "tag bleeding" nightshade targets niche concepts, so if you have 1M images to train and your target concept is just 0.001% of that (so 10 images), you only need say 2-3 images to completely throw off this concept, if you were unlucky enough to land enough epochs on tampered images, on the tampered timesteps HOWEVER popular concepts are significantly harder to fuck with, if your target concept has at least 50 images, good luck trying to throw it off lol
>>29929 so it works like normal tag poisoning, got it not something to worry about then except for e621 nerds lol
>>29929 >basically it relies on so-called "tag bleeding" So basically it's gonna be next to useless on anything but CLIP? What's stopping me from giving it the lightest denoising pass with the shittiest denoising model I can find and completely bypassing their garbage?
>>29931 If a model uses a different text encoder and VAE than the ones they target, I think it should be completely immune, other than the image looking a bit noisier. And I think they targeted SD1.5?
>>29930 ironically you can try to fight it by using seemingly false positives on tagger models, these tags aren't necessarily "false", they may be detecting shit prompts that got encoded into the image by nightshade >>29931 tag bleeding is an achilles heel of any model that uses tokenizer (all llms) but regardless, it's actually more like "semantic tag bleeding" so you can actually bleed "vanilla_(nekopara)" into "cat_girl" however nightshade specifically relies on clip to provide false guidance so you'll be effectively distilling your model off of clip >What's stopping me from giving it the lightest denoising pass with the shittiest denoising model I can find and completely bypassing their garbage? you don't have to bother to do that, this shit will not work even if you sit doing absolutely nothing there's a good chance newer models like 4o that don't rely on diffusion aren't affected in the first place, it'll just have the same effect as garbage captions would have
>>29933 >you don't have to bother to do that, this shit will not work even if you sit doing absolutely nothing So hold on, they created a poisoning method that does absolutely nothing in the end and they're jerking each other off over it anyway?
>>29934 That's all those anti-AI techniques in a nutshell. It's just a second level grift.
>>29935 funny they're even more behind than local development
>>29934 it does work under very specific conditions and these conditions are generally unachievable unless you decide to train your model sd 1.5 way on a shitton of images poisoned in the same exact way >they're jerking each other off over it if you're talking about artfags, absolutely yes i don't think there's a serious discussion about it, even a relatively minor paper such as minsnr got twice as much citations
>>29927 >had less cultural impact What cultural impact does your life have, 8non?
>>29926 you're part of the problem
>>29940 plantmilk oat had a serious artifacting issue unless you used karras scheduler, so I didn't even touch the later versions oh but nta
>>29941 >cartoons had more sexual impact than anime
>the genshin impact on my dick and wallet
>>29941 Yes hentai is unpopular weeaboo garbage and not even the average gooner wants to look at it.
>>29946 then explain why Princess Peach is so high in the rank
>>29947 nintenddogshit is clearly ACO in nature
nintendo is a japanese company thoughbeit
>>29947 Almost all Peach art leans more towards cartoony rather than anime. And that's a good thing.
Then everything changed when the Irony Nation attacked.
>>29941 cob won
the fuck is a cobson?
>>29953 like dobson but weaker
>reforgeui >[DISCONTINUED] what?
>>29953 wojak variation
>>29957 thats a childs
>>29958 Is this childs in the room with us right now?
Stop shitposting.
(1.82 MB 1024x1024 ComfyUI_39018_.png)

>>29957 eh, sure
what is your guys' go to base model nowadays? i'm honestly finding i liked illus 0.1 and noob-vpred06 more than noob-vpred10 nowadays
(2.60 MB 1440x1600 ComfyUI_39040_.png)

>>29963 Lately I've gone back to using v-pred1.0 without any merges. Earlier v-preds look a bit nicer but I miss the extra knowledge, lot of styles/chars I use are around 100 pics so it makes a big difference. Also if I bake a shitty style on the earlier v-preds then use it on 1.0, the shittyness seems to get diluted much faster than the style itself.
>>29964 yeah it's having to trade that nicer look for the extra knowledge that really bugs me. i wasn't around for when 0.75s launched so i'll have to try it out and see how it compares
>>29963 I stick to 1.0 now (and occasionally 102d final), older vpred models had even more issues with schizo contrast compared to 1.0, also less knowledge as anon mentioned above 1.0's base look is a bit unpleasant but I find it to be fixable by using some lora you like at low weight.
>>29963 I'm still using noobai-XL-Vpred-0.75-cyberfix but I should spend some time and test some other models properly.
weirdly calm today ah, it's Easter
What's v-pred1.0? also, fuck this captcha
(3.74 MB 2560x1536 00190-3170755564.png)

>>29969 noob vpred 1.0? a disgrace and the last version of that bake
>>29970 sad it's still the best
>>29971 1.0 2D is best
>>29971 as long as you use a lora and controlnet, yeah it is
>>29972 i mean true but that's not a base model
>>29974 Since the cope for base 1.0 vpred is "just add a lora", and 102d is essentially just the base model with a few loras shitmixed in, it is the best version.
>>29975 >a few loras shitmixed in Have you read the recipe? That's fuckload of loras.
>>29975 Except for not being able to pick which loras to mix in. Works better if you pick one that matches the style you're going for. But yeah it's fine for general use.
>>29977 what's the recipe? I never actually checked arca about it because I can't read farside moonrunes.
>>29979 10% luck 20% skill 15% concentrated power of will
>>29980 and 100% reason to overfit the checkpoint
Why is 8chan catbox script limited to /hdg/ only? Shouldn't it be sitewide?
>>29979 This is on the civit model page. I assume first pic is the recipe of his first model, the description is kinda convoluted
>>29983 Wait, it's on civitai now? What's the checkpoint name? Also jesus christ that is a lot of LoRAs, crazy how it manages to feel like the stablest/least noisy shitmix of vpred 1.0.
>>29985 which 102d version is the best? is BO worth using? I've been on ΣΙΗ but sometimes it looks too artificial for me
(3.43 MB 1664x2304 00187-2767487591.png)

>>29986 NTA but final or custom are fine
>>29987 thank you bwo
>>29982 The script adds loading time to an already slow webpage, also I want to avoid conflicts with existing userscripts Is there any other new refugee board besides /slop/?
>>29990 I mean there are a bunch of ai threads all over the place.
>>29963 https://civitai.com/models/943557/csl-noobxl I feel like this smoothes out some of noob's default shittiness nicely without destroying most of the things it's good at. I still use loras a lot though. Illustrious is unusable to me because it gives garbage results on 1.0 CFG and I refuse to go back to using negative prompt after realizing this is twice as fast. On most models it actually looks better like this but illustrious uniquely looks awful without a negative prompt for some reason.
you ever feel like you've burned through everything you've wanted to gen? im stuck 1girl genning now and no new ideas flow
(677.03 KB 657x910 1790347126580117.PNG)

This is how I read you.
(814.59 KB 1920x800 00093-4286512340_cw.png)

Gen something cool and funny.
(3.95 MB 2688x1536 00122-844704649.png)

(3.61 MB 2688x1536 00082-2112874541.png)

>>29993 Not really, in fact, even if I'm been on this hobby since the 1.5 days I keep getting ideas about things I want to gen, there is a lot of smut that I haven't explored yet >>29994 Why are you a little anime girl
>>29989 There's /trash/ with the /slop/ thread. Also /ais/, but I think it's more chatbot focused.
>>29993 Yes, I'm back to doing a few gens whenever I see a cool new concept, but usually I can replicate it quickly and I'm back to the start again. Now I'm between making my second Lora or resuming my pixiv again, and I kinda don't want to do the latter because it feels more like a job than a hobby.
>>29999 >and I kinda don't want to do the latter because it feels more like a job than a hobby. same, it really feels like pressure to me despite pixiv slop performance being completely inconsequential kek
>>29993 Often, but I just have to take a break for 2-3 days to come up with new stuff.
Think what I miss most is people asking for tech advice, helping them made learning all this crap feel worth it. Not like I'm ever gonna use ipadapter or regional prompter when genning for myself, I don't even inpaint fingers unless it's for a public post.
>>30002 >I don't even inpaint fingers unless it's for a public post. Really? I properly inpaint gens that I like even if I don't plan on share it anywhere
(1.59 MB 1024x1024 ComfyUI_39310_.png)

>>29607 Why is it not listed?
(1.08 MB 1024x1024 ComfyUI_39540_.png)

Still amazed at what the AI comes up with sometimes. Sure, the mother is the inexperienced one.
Oh hey, you guys are here.
never heard of that 2d model you guys were talking about so i tried it out along with some other vpred models dont think my shitmixes did that well this round but they weren't the worst at least i dont want that 2.5d style so chucks and jankun are getting thrown out, they both look good with the miku but you can see the style drift with the opossumachine gura i can see why you guys like the 2d models they seem pretty well-rounded and a good, reasonable choice i dont think ive ever managed to get rouwei to look good one of the tags was "kneeling on liquid" and only ch and lhc got that right i think ch and r3mix are my favorites this round >>29993 good way to get out of that is to just do other shit for a bit i used to gen for quite a while and then i got bored of it so i switched to making bots and im on a creative block with that now so im back to genning
thats a children
nice
>>30011 that's a thousand year old shark trying to get me arrested on bestiality charges actually
>>29991 >>29998 Enabled the script on all boards, let's see if it will cause any problem
>.webp reeeeeeeee
>>29985 damn esl is strong with this one. >The custom version is a model based on the previous model of the final50 version, where the weight ratio of the NAI-XL original model was increased to 65, followed by further adjustments. It is not guaranteed that the model will produce more aesthetic images. >While the final50 model is designed to generate highly 2D-flat images, this model has been tuned to produce a more standard like. Please choose the model that best fits your taste. It is likely that the final50 model offers a better LoRA response. so I guess "custom" is slightly closer to base compared to final?
>first time someone tipped me buzz >it's a vore+digestion fag civitai was a mistake
Astra posted new V7 samples, but they're all ponies, no hands.
>>30018 I wonder if it's at least better at ponies than v6 was
(1.80 MB 768x1280 Clara(A) (01).png)

So uhhhh HiDream, what's that about?
>>30022 >achieves state-of-the-art image generation quality within seconds yeah, assuming you have 27gb vram looks like a Flux sidegrade at best, nothingburger until someone finetunes it on booru
>>30023 Reddit claims that it doesn't have the training issues of flux
>>30024 Stability also claimed SD3 was trainable, and it turns out it's not that simple. Don't trust that shit until you see material proof.
>>30024 I don't think it's really possible, even if it's not distilled and properly trainable. We had to wait for almost 1.5 years for something like noob to come out and it's just xl which has some long inheritance history - kohaku, angel and finally lax's team. Tuning something bigger will probably require alot more gpu time. Not even mentioning the post usability of that, like if it has problems to even just run in full precision on consumer cards and knowing how important loras are, train them will require even more vram, it might just don't get much of popularity, or end up in some distilled 4bit sameface/style hell when all loras would look the same because no one will be able to bake it in a proper precision
>>30028 >knowing how important loras are, train them will require even more vram, it might just don't get much of popularity if the model has proper dataset, loras are not that important. if you really need a lora for that no name girl from some 1997 anime and are a vramlet, training on some saas platform or renting a cloud gpu is possible
>Trained on NoobAI-XL (NAI-XL) Epsilon-pred 1.1-Version fucking retarded
So anyone tried illustrious 2.0? Is it better than 1.1 at least? DL speed is very slow for me right now, should I even bother?
>>30029 Can't agree with that. Any proper baked lora are improving any xl model, even using with a low weight to stabilize the model is somewhat good. Wonder if it's still true for v3, will probably never know. Might be not true for newer models anyway. And what about somewhat middle 12gb tier? Cards with so insufficient vram won't be able to inference it even at 8bit or q8 quants
>>30031 >should I even bother? Not really, if you are using any vpred model already, the only illust version that could potentially be better than any other model we have is illust 3.5 vpred but we are never going to get it lmao
>>30033 I expect they'll switch to another business model once the donation pool dries up. Still who knows whether they even managed to surpass v-pred 1.0 in terms of anime, and it's missing e621(?) plus six months of loras.
(3.77 MB 1536x2560 00257-3238939048.png)

>>30034 >Still who knows whether they even managed to surpass v-pred 1.0 in terms of anime, and it's missing e621(?) plus six months of loras I'll still hold my hope up even is a little pointless at this point, I would miss e621 wild poses tags but I'll gladly trade them for a better mostly pure anime base model
>>30029 >if the model has proper dataset None of them do and none of them ever will anymore because """ETHICS"""
>>30035 long leg
(4.80 MB 1536x2560 00293-3254376940.png)

>>30037 I think is fine but if you have a clear reference on how it should be, I'm all in to fix it
Does Danbooru keep a history of tag changes somewhere? I was trying to get Sue Storm at some 670 pics, it can't even get her hair color right. So I assume it must've been renamed, last I had this with hagakure_toru as she was still "tooru" in the noob dataset.
(3.01 MB 1536x1920 20250421120234-4252354669.png)

>>29817 imo 102d_custom looks much better than r3mix. 102 is in the mix for r3, test them side by side and see how identical they look. but you will notice 102dcustom has less AI slop look without the need for negative loras.
>>30043 Must be something else then in this case, I saw those aliases on the tag wiki page and tried them already. But thank you, still will be useful in general.
(2.12 MB 1248x1824 catbox_6n54kl.png)

(1.55 MB 1248x1824 catbox_ng8voh.png)

(1.75 MB 1248x1824 catbox_xrl5m5.png)

(2.71 MB 1248x1824 catbox_7729jp.png)

Just some testing, this reminds me of AOM/Based64 though.
>>30018 Good news, style clusters are finally starting to work!
>>30047 Uohhhhhhhhh! !! !! Child erotic !!!!!! Child's belly and chest !!!!!!
>>30049 uoh no
(2.14 MB 1824x1248 catbox_ndl8v6.png)

>>30049 uoh no
>>30057 size was raised to 15mb a while ago because some anon asked
>>30058 the original png for the one i converted into webp was 115 MB
>>30058 some anon should ask about not pruning metadata, at least on the AI boards
(3.79 MB 1664x2432 ComfyUI_39783_.png)

I had to write a relatively short guide for something (not AI-related) on rentry and it was fucking painful, props to the people who do this shit regularly for AI stuff.
>>30062 Yeah writing guides sucks, especially for tech. Everything changes so quickly so you have to update often or just abandon it eventually and it's impossible to gauge how tech literate your readers will be so you're always either dumbing it down too much, making it too advanced or both.
>>30063 >Everything changes so quickly so you have to update often or just abandon it eventually and it's impossible to gauge how tech literate your readers will be so you're always either dumbing it down too much, making it too advanced or both. Nah in this case I what I had to explain was relatively simple, it's just that there were *many* steps and I probably made the mistake of prettying it all up (or trying to) before I was done writing it.
>>30064 The "make it look pretty" curse falls on me too. I usually write everything out in notepad first and then copy it out to rentry to add formatting and images to avoid that.
Anyone know of SDXL VAEs that improve over vanilla? I have lpips_avgpool_e4.safetensors and sdxlVaeAnimeTest_alpha67500.safetensors but I mostly just use the former.
>>30067 Improve what?
>>30059 ok yeah 115mb is too high I think the site highest is 50mb >>30060 That's an issue for the /site/ devs rather than for the board owner. I don't think they want to do that though due to the possibility of abuse.
>>30045 >looks much better i'm not stating that either r3 or 102d is better looking, but 102d is definitely the more boring one to use. output composition is so uninspiring. too bad you can't get the same schizo creativity base model got with anatomy stability of a shitmix on later timesteps.
is this where the /h/ /hdg/ refugees congregate now?
>>30071 Unfortunately.
>>30072 seems too slow for that, you sure?
>>30071 no, we're from /b/
>>30073 be the change you want to be and proompt
>>30060 All you have to do is post webp This site doesn't prune metadata from webp >>29793 >>29713
>>30074 Are we? I thought almost everyone here was from /hdg/
Something in the recommended quality tags of noob v102d is preventing it from genning dark scenes
>>30079 >I can't "coom" unless I see dark!!1
>>30080 fuck off, /hdg/ troglodyte
Go back to treechan
(2.59 MB 2688x1536 00000-3953224458.png)

(3.02 MB 1664x2432 00040-3872013839.png)

>>30079 Don't use any quality tag or snake oil on 102d and only use neg tags if you really need them
>>30083 >Don't use any quality tag or snake oil on 102d and only use neg tags if you really need them déjà vu i've just read this exact? line before
>>30083 >clothes
(1.88 MB 1664x2432 00147-3859584377.png)

>>30084 :^) >>30085 >tasteless nude
>>30086 Why is it darkless?
daniel stevens (owner of the once-fun-but-now-full-of-gabniggers-and-boomers fediverse instance poa.st, c*n*di*n cuck extraordinaire and allegedly an ex? halfchan mod/admin) wants to buy halfchan for 1 million usd, if it goes through you can say goodbye to lolis
>>30088 >cuckchan who cares at this point
>>30089 me because i hope that the tourists will move back asap
>>30088 Isn't halfchan fucked beyond any saving now? I get the sentiment, it's always bad when a part of the old internet dies, but is there even saving it?
>>30088 cuckchan didnt even allow lolis outside of /b/ anyway
>>30091 im certain it's dead, this is like moving the rock, and all the insects scatter and hide elsewhere
>>30084 And you didn't listen? Anyway, if not for general use then at least keep it in mind as something to try, when the model refuses to do your style/composition/bodytype properly. "very awa" might as well be "score_8_up" and other quality tags have their own bias too.
>>30091 once halfchan goes back up (and sadly it will go back up), everyone is going to flock back even if it's required to suck new admin's dick to post there. this week felt at least a little bit like old internet though, I missed the chaos.
>>30093 holy shit he is pogging
>>30096 It's not really sad that it will go back. Unfortunately, "moderation" is the way of the world now, and everywhere will require it. The Wild West days are mostly over. Having 4chan as the poster boy of the "rebel internet" allows all other chans to operate away from the spotlight.
got my first Civitai TOS violation today apparently the /hdg/ loli rules were still pretty tame
>>30098 >everyone (site owners) needs to abide by the modern web rules because uhm... they just need to, ok?! the only winning move is not to play (read: give a fuck)
>>30100 and then suddenly your board gets spammed with cp and abuse letters are sent to the hosting
>>30101 i didn't say "just let people post cp lol"
>>30102 nah I meant if your board doesn't conform to glowie rules they are just going to spam it with cp as they usually do
>>30100 there is also a legit threat to getting deplatformed, because all the major hosts choke on the big ESG cock. (((harmful users))) will just make trying to run your site more expensive, this isn't about following some "modern rules" because everyone else is doing it
>>30074 where did the /h/ /hdg/ refugees go then?
>>30104 you can mitigate this by just blocking all vpn ips, third world ips and obviously kike ips
>>30057 What are these models? Custom mix?
>>30106 holy shit, this guy has the solution. brother, go buy halfchan and uncuck it for us, you're our only hope
>>30108 wallow in your irony, this has literally never been tried and i'm 100% confident it would be the solution
>>30105 No one knows.
>>30105 just here and treechan afaik
>>30112 sexy, is this ΣΙΗ?
>>30112 hand is a bit wonky but i REALLY like this mix
hmm v4 vt is surprisingly alright
>>30117 <touch the cow
I wonder what happened to /d/. Are they also here with us, or did they just die off? I know /aco/ is dead for sure, what a waste.
>>30121 I think a lot of the slower boards will be super dead when they come back, and I don't just mean the diffusion generals. I hope at least the /hdg/ schizo has a psychotic breakdown from this and disappears so we can pick up the pieces, at least.
>>30122 I think there were too many styles of shitposting for it to have been only a single person.
>>30123 Makes sense. I hope they all have psychotic breakdowns and go post on /x/ or something.
>>30123 Nah, it was definitely like 1 or 2 guys the shitposting completely stopped when it was their bedtime
>>30121 maybe just posting in the slop and degen threads?
Newbie here, I've been playing around with dmd2 and regional prompter. When using both, the regional prompter doesn't seem to work (I assume because my cfg is at 1). Does anyone know if there's a way to make it work while keeping cfg at 1?
>>30081 "/hdg/ trogdolytes" are the only people caring about muh dark colors muh chiaroscuro muh glowing abstract shapes under the moonlight
>>30122 seems like he's here too anyway
(952.68 KB 896x1024 tmp0u5aphnw.png)

(881.50 KB 896x1024 tmpwov5pi05.png)

>>29732 what about chibis? are they the superior body type or are they too far over the line?
>>30131 sex with cihibs, holding their body like you would an onahole
>>30131 chibiboxes?
Chibi? always in negatives
>>30134 this but "chibi inset"
>>30122 I fucking hope 4chan never comes back. Did you use anything aside from hdg? Every fucking thread and general is infested with schizos and power tripping troon (actual fucking troons) mods. It's not just /h/, the whole website is garbage nowadays. Maybe the hack and jannies leaving would force them to change their approach to moderating the website, but I doubt it.
>>30136 >Did you use anything aside from hdg? NTA but yes, I used to visit another boards and I'm glad most of good posters on those already migrated here, we still have a few schizos but thanks to the thread specific IDs is really easy to spot them and mods actually do their job here
>>30136 nta either but: /aco/ realism thread arknights drawpile on /vg/ self-improvement general on /fit/ all of /vst/ lot of comfy places I'm going to miss
>>30136 Just the porn for me personally, but theres a bajillion places to fill that niche on the internet.
(869.50 KB 1024x1024 ComfyUI_39973_.png)

>>30134 Wouldn't you just become chibi yourself? It's an artstyle not a bodytype.
>>30140 That is just a minigirl.
>>30136 Why would you not want it to come back though? A lot of those people will migrate if it doesn't, and even if they get moderated here, it's better if they just never find a reason to come here in the first place.
>>30141 head to body size ratio is same as >>30131 I thought that's the main condition minigirl has human proportions and is just smaller overall
>>30143 is it? The torso seems larger than the head.
>>30144 eh, alright maybe it's halfway
>>30133 https://files.catbox.moe/xpsgpr.png https://files.catbox.moe/fenl96.png >>30135 Sadly chibi inset doesn't work that well for me. I keep encountering the "strange creature" phenomenon regardless.
>>30146 Try this: chibi, chibi inset, huge head, SD character , doll, crowd, multiple people, group, deformed, mascot, stuffed animal, dwarf. Adjust weight accordingly. Works much better for me. You can also try to make more detailed prompt for the location/background so it can fill the area without inserting random shit.
uoh cunny cunny plap >>30107 yup
>meets the donation goal >no model ????
>>30149 oh, oh nononono... see, the individual models' bars are for... funding the individual models' training. the global bar is for RELEASING them. the gooks can and will try to scam you twice in a row.
>>30149 hory shet, real?
>>30140 I can feel an anal fissure just from looking at that pic.
>>29992 Would you mind sharing a catbox or some general advice? For me it's the opposite, I got some ok results on certain illustrious models but having a hard time getting it to work properly on noob/pred models.
(1.99 MB 1600x1200 catbox_pl6qs4.png)

(1.99 MB 1600x1200 catbox_yozcpu.png)

(1.97 MB 1600x1200 catbox_i1cjcr.png)

thats a kid
>>30156 (adult)
(2.24 MB 1568x2016 00102-2567771583.png)

>>30158 based
why are you giving money to kurumuz?
>>30014 anyone else getting this? if there's no metadata right clicking the icon again won't make the box go away.
>>30136 I like board rules much more here /h/ for example, no nonsense like banned tentacles and loli, /d/ fetishes contained to separate threads instead of making separate board, can posf /e/ freely. 4cuck /h/ board rules are pure nonsense
(1.74 MB 1568x2016 catbox_ypf1ff.png)

>>30164 good gen
>>30152 she's made of onahole rubber
>>30136 had been off 4chan for years until the nai leaks happened, and lewd SD really took off from hdg brought me back, but only to hdg. the enthusiasm and information sharing really reminded me of the good 'ol days of 4chan, actually. hdg threads were maxing out twice or three times a day. at the beginning anons were happy to teach and help new people. every week something new would drop and revolutionize what everyone was doing. you had to dig through threads for info, it was super engaging and electric... that was until civitai popped up and centralized everything about open source diffusion models. then all the help and energy for the sake of this new technology disappeared almost overnight. eventually the SaaS vs Local flame wars pushed me away from hdg, and i haven't looked back. eventually it became just like the rest of 4chan, full of schizos in a pissing match thinking they're hot shit. so dicks out for 4chan. it had its glory days but those died out long ago when moot sold the site. soyjacks did us a favor by taking it out back.
Hope 4chan comes back soon, can't stand this lawless place
>>30169 Lawless in what form?
>>30170 it's a shitpost
>>30170 People are posting things he doesn't like. I actually would like it if IDs were enabled here.
>>30172 >IDs At the very least it would make the life of the "why so blurry bruh" posters more difficult so it may be worth it.
>>30173 rent free
This place was really nice before the half chan disaster. I just hope it can recover after this is over.
>>30174 >its afraid
>>30175 Without jannies protecting them, shitposters don't have much power. Just ignore them and keep posting nice pictures. Trolls leave if you don't engage with them. Thread IDs would be nice, although I feel like they're not as effective as people seem to think.
>>30175 you filled a grand total of 12 threads in 2 years. not exactly hard to have a nice thread when you're the only one posting
>>30177 /aids/ has become much more pleasant to read by having IDs
>>30155 aoi yuuki (real)
(2.49 MB 1920x1536 20250423052222-2348352428.png)

>>30161 Why are you shaking?
Jordach is blogposting in the treech thread. Says he became a pedo recluse for six months, but he's better now and he's going to "fix" noobAI.
>>30168 >the enthusiasm and information sharing really reminded me of the good 'ol days of 4chan, actually Yep, I had a similar feel. I was a bit late to the party but it still was great before schizos took over.
>>30184 I got the same feeling in the first month or two of Pony, even if oldfags say things were already bad by then.
/hdg/ was always great. Some people are just thin-skinned.
>>30185 tag mining was like giving meth to an alzheimers patient. it worked great once but the general still kept declining fast.
>>30186 this, in India we are used to everything being smeared in shit, westerners are just way too sensitive to our culture.
>>30187 we baked a shitton of loras tho the lora rentry almost felt like a community project, and Civitai lagged way behind
avatarfags ruined /hdg/
>>30186 Safe to say its quality steadily declined over the general's lifetime. Depends where you draw the line where "great" ends.
>>30190 I'm still mad they never gave me a name
>>30162 The script treats "no metadata" as an error, on error right click will retry the load & parse process, which I agree is both counter intuitive and wrong behavior: there's nothing to retry in this case Updated the script, now right click in absense of metadata will hide the message container
>>30194 it's "absence"
>>30186 It really took a nosedive once more people found out about it lmao
>>30196 yeah pony blowing up was the final nail in the coffin
Notice when a Civit lora is posted for both illustrious and pony at the same time, the illu one gets almost exclusively just downloads while the pony one gets mostly onsite gens.
>I know you were looking at me, senpai
>>30010 I find it much more difficult to compare SDXL models than SD1.5 merges as now most of them are finetunes, the extra bit of training put into them more often than not changes how they react to quality tags / samplers / schedulers / etc. An X/Y plot may not be a fair comparison between for example noob and rouwei because they have different recommended samplers, steps, quality tags and so on But I'm also too lazy to spin up a convoluted comfy workflow to do it correctly so now I use whatever model I found acceptable in rotation
>>30201 >so now I use whatever model I found acceptable in rotation This is the way. Even in 1.5 days, there were models that were better and worse at certain things. Just use whatever gets you the particularly best image you want.
(2.21 MB 1824x1248 catbox_v59q5q.png)

(2.49 MB 1824x1248 catbox_r4s1hb.png)

(1.76 MB 1248x1824 catbox_6v4t5z.png)

(248.55 KB 559x277 uohno.png)

More testing and this game better be good
more Civitai cuck rules https://civitai.com/articles/13632 >no incest >no urination >not drunk, drugged, under hypnosis or mind control also >Require all mature content to include generation metadata
>>30207 img2img with any nfsw tags or autodetected as nfsw are locked to a minimum 0.5 denoise Well it was only a matter of time, enshittification for the sake of profit.
love how they ban this tame shit, but bara-hyper-nipple-dick-growth.safetensors gets to stay
>>30209 That's just typical redditor "morality".
>>30210 Yes it's really just responding to what moralfags complain about. I blame Civit for caving in at all, not for which topics were chosen.
>>30204 Is that rpgmaker ero skyrim?
>>30212 I hope not, I think TES is shit. >>30213 oh no
>>30214 >oh no, it's that kind of game or >oh no, they found out Anyway there's a surprising number of games like that now, just pretending to be sfw to get on steam but a complete joke if you try to play them without the uncensor patch.
>>30161 Because I want the best and local efforts suck.
>>30215 Well I knew what kind of game it is and I made sure to get it on my 2nd steam account. The dude that posted that review is clearly a lolicon in denial.
>>30207 >Diapers Isn't that like what most of their userbase is genning kek
>>30213 >It's decent, but devs are based. Downloading it right now, thanks for the recommendation!
>>30217 \rh, could be furry, or an actual pedo, or something else entirely. No reason to stay what you hate when it's a choice.
>>30218 That's the meme, but I don't actually remember seeing any on Civitai. There's a lot of furry, the usual nyantcha blacked stuff and a huge amount of photorealism, more than half of all nsfw for sure.
>>30217 >I made sure to get it on my 2nd steam account. Absolutely no balls.
>>30222 Just buy it on their site. Save them paying Steam's 30% commission, and less incentive for them to keep wasting money and effort on these fake SFW versions.
>>30222 >no balls. Yes, but I do have a few co-workers on my friend list.
>>30224 you can mark games as private so they don't show up for your friends, including achievements and stuff still a good idea to not have the game associated with your credit card info, you used a gift card right?
>>30127 Sorry nobody answered anon, but I don't think anybody here uses dmd2. Try other methods such as Lightning LoRA. And make sure you're on a speedy UI like ReForge.
>>30225 Yeah I know about marking games as private but it still doesn't feel "safe" and I did use my credit card, oh no no no.
>>30225 >still a good idea to not have the game associated with your credit card info still the same old cowardice, huh
>>30223 How do you buy it on their site? I only see a steam button.
(3.60 MB 1792x2304 00067-2286126074.png)

(3.57 MB 1792x2304 00072-3665371823.png)

(3.13 MB 1792x2304 00019-33602415.png)

(2.76 MB 1792x2304 00000-3382075523.png)

my local wife.. Iazy slop but, haven't used this shit in ages and its still as much of a pain in the ass as I remember it being
>>30230 skill issue probably
>>30207 >banning noncon lol
>>30235 her bush should be the size of the amazon forest
does anyone use forge classic or base forge instead of reforge? https://github.com/Haoming02/sd-webui-forge-classic https://github.com/lllyasviel/stable-diffusion-webui-forge ive been thinking of moving to one of them because of development stopping on reforge
>>30238 I you're worried about staying up to date, the only "safe" option is comfy right now. There was that swarm UI thing that is a layer on top of comfy for ease of use, maybe that makes sense to you?
>>30239 i also use comfy but i prefer a1111-based webuis for some things i think im going to try out the new forge version because of the improved inpainting interface
Just use Invoke
>>30242 Ah yes. Invoke will for sure have cutting edge features implemented not too much later than everything else.
>>30242 Hey, that's my shtick.
>>30243 What "cutting edge features"? All the snake oil extensions everybody drops a week later?
>>30245 Or all the nice samplers that nobody uses, yes. But if you don't care about those, then why would you not use forge / reforge anyway?
>>30243 >>30246 nta but Invoke is pretty cutting edge. The (regional) inpainting features and resizable+infinite canvas are pretty fucking huge QoL features.
>>30246 I am using reforge because there still are some things that keep me here. But Invoke is looking more and more like the superior interface for an inpaint and sketchfag like me.
>>30237 she read shaven girls are more popular
uploads are really laggy here now
>>30252 whenever i post with no attachments i need to wait like a minute before it actually goes through
>>30252 the website is very laggy in general, can't handle the refugees
>>30252 speaking of, is .se dead for everyone else?
(1.51 MB 1120x1280 ComfyUI_40385_.png)

>>30254 sounds like a great excuse to post gens
>>30257 meaty oni
>>30207 apparently its pressure from payment processors like usual what i dont understand is why they dont call them out specifically to pressure them back, and instead choose to be the ones who the users will be mad at
>>30262 anon's cute wife...
>>30207 Civitai gets worse and worse every day Can't have anything nice but we can definitely have 50 pages of bara and 100 pages of furry Hope someone makes an actually good site soon
>>30153 dmd2 4 steps lora https://huggingface.co/tianweiy/DMD2/tree/main illustrious/noob models I always have to do 6 steps or it looks like shit. 1.0 CFG (don't worry about your negative, 1.0 CFG disables it). LCM sampler, Simple scheduler. Use good artist tags, certain artists have an incredibly strong effect, sometimes a bad one. Experiment and mix the best ones. 70% of getting a good pic is artist tags in my experience. Sometimes I just remove all artist tags on a recent prompt out of curiosity and I'm appalled at how godawful, shitty, and generic the picture looks. This is the fabled "ai slop" that everyone hates, gens with no artist tags.
>>30248 I haven't used Invoke but the Krita extension suits my use case personally, especially since I can do custom workflow shit with it.
>>30266 >Krita extension this shit filtered me real hard
>>30264 >Can't have anything nice Rape, shit, sex with toddlers, piss, and explicit deepfakes are not "nice". Your mind is warped beyond repair if you think publishing that stuff is somehow better than images of a naked bear dude. And if people didn't flood every place with their morally reprehensible Clintonciding fetish, normies wouldn't be crying for bans in the first place. Keep it to yourself and don't force your pedo pride parades onto other people.
>>30269 Kill yourself.
>>30270 t. groomer pedo
>>30200 Catbox please?
>>30275 everywhere i go i see her face...
>>30264 Well mainly there's just no reason to use it, other than as a way to archive and share loras/checkpoints. So there restrictive rules barely even matter. The social aspect is limited to four emojis, very few people even use comments. You're better off posting gens on R34/pixiv and redirecting people to twitter where there's a culture of interaction. And the whole idea of being able to cash out buzz into real money is bait. The only way to earn yellow is when other users tip you their yellows; and bounties. Bounties are basically open commissions where only one person gets paid in the end, so why wouldn't you do regular commissions instead.
>>30277 this the users themselves make it dogshit for using for image sharing >buzz pls babi reedem 4 me >that one retard who uploads every single variation of the same exact prompt with maybe a pixel that's slightly different x50 >furshit everywhere
>>30278 You can blacklist furshit, and diapers, bara, etc. I agree with the rest.
>>30281 I do not understand the use case for multi-style loras
>loras What's the use-case?
>>30240 Box?
Man, this multi image thingie is pretty nice. I wish we had that a long time ago.
>>30284 https://files.catbox.moe/yk5i2g.png almost forgot about you. Sorry, buddy.
(2.65 MB 1328x1776 00018-1608120676.png)

(2.68 MB 1328x1776 00007-3750725381.png)

(2.45 MB 1328x1776 00022-1824440239.png)

>>30163 >can post /e/ freely yeah but poor /edg/ is going to die off completely at this rate
>>30288 Good. It should have never existed.
>>30288 Surely >>>/slop/ must be a very healthy board, surely.
>>30288 /edg/ was the better /hdg/, where people posted images and discussed tricks and there were no schizos. If we get rid of shit posting schizos and the thread does not move too fast, I don't see a problem with having /edg/ merged in. /edg/ can split out if we ever reach critical mass and there's too many people posting at the same time.
/edg/ was utter slop and the only thing they discussed was how to best suck highlightfags cock
>>30292 >still seething
>>30292 This guy is still angry at highlights. Incredible.
Now that it's not here, I miss highlights bros
>>30297 what's wrong with his dick?
>>30298 I don't know, you tell me. What's up with it?
>>30299 >What's up with it? necrosis
>>30300 Ah, you see the darker skin color of the dick and become confused. It's fine, I don't think it looks weird, but thanks for bringing it up.
>>30301 i know it can be normal for dicks to be slightly darker than your skin color but the dude's dick's about to fall off
>>30302 I really don't see it, I also think the lighting would lead to something looking like that. But maybe I'm colorblind and can't see shit or my monitor is too cheap.
>>30303 nta but I think it's cause there's no diffuse lighting given how shiny the guy's torso is, the dick should look (very rougly) like this
What were the specifics of training on v-pred again? Enable v-pred, z-snr, scale vpred loss, and disable noise offset. Anything else like scale weight norms, IP noise gamma, scheduler/optimizer arguments? Also I'm still running AdamW, if anyone has a better idea.
>>30257 onibox?
>>30304 ah, that makes sense. I guess i can work with that.
>>30305 >scale vpred loss Don't https://github.com/kohya-ss/sd-scripts/pull/934 It has been unnecessary for more than a full year now
(763.02 KB 1600x1216 GpSEaE4bYAEVJTx.jpg)

(356.44 KB 832x1216 GpSZ_s_bYAUKHyY.png)

(405.57 KB 1472x1472 GpR1uKRbAAAtvmf.jpg)

(390.63 KB 1024x1536 GpR8ewnbYAIx_uZ.jpg)

Some V4.5 posts from twatter. The classroom doesn't look too bad if you ignore the poor bloke with the split head.
What do you think she means by this?
>>30310 Nothing. It's just a greeting modern women do. Don't read too much into it, creep.
>>30311 Oh shit, sorry, I didn't mean to oppress anyone. I guess this is also a normal way to greet new acquaintances, right?
>>30309 What's up with the glow around characters? Every image model I've ever tried has this to some degree but I feel like it's 10x worse than usual on V4 and seemingly v4.5 now.
>>30313 It's a thing many artists do to make the character pop more from the background.
>>30314 I wish they wouldn't, it looks so bad.
>>30314 >>30313 It's not to make it pop, it's because the chars are drawn, usually, by the mangaka, while the army of assistants will draw the backgrounds. They have to be pasted on top, so you see the white border on them to mitigate the fact that the characters are usually not lit correctly, etc.
>>30316 Ironically I think it ends up making them look even more pasted-on
>>30317 Yeah, it does. I don't think they care much, it's just a thing that happens since the old days, when you would cut out the characters and physically paste them on top of the backgrounds. Likely these days they can do this digitally, and this is why not everybody does this.
>>30316 think he meant the white glow in pics 1 and 4, not the extra white outline in pic 2
>>30319 That's a remnant. Nowadays you do this digitally, so making the extra outline a gradient makes it look less obvious.
>>30319 Yeah that's what I meant.
>>30306 Sure. I had a little help with controlnet, 0.2 anytest. This is same seed without. https://litter.catbox.moe/0gm1wi.png Checkpoint is ΣΙΗ, metadata garbles the greek letters. And style lora from https://archived.moe/h/thread/8405374/#q8407916 I'll eventually rebake and put it up on Civitai, just gotta inpaint like two hundred hands/eyes because the dataset is fucked.
>>30309 localsisters... not like this...
(1.03 MB 1280x768 ComfyUI_886_.png)

>>30323 Okay but can NAI do this?
>>30281 im surprised they made one considering how i only post sporadically in hdg threads
What are reforge users going to use now that there's no more updates? Back to forge?
>>30326 stay on one commit for a year until there's a reason to switch to something else
>>30326 gradio 3 already had major security vulns but its not like SDXL itself is going to somehow be updated so just keep using it for now? do NOT switch to regular forge, it has obfuscated code because illyasviel had a meltdown over not properly attributing bits he took from comfy. forge classic is an option if you absolutely MUST be on something that's actively maintained. or comfy lmao
>>30328 >do NOT switch to regular forge, it has obfuscated code because illyasviel had a meltdown over not properly attributing bits he took from comfy I merged the PR to remove this several months ago btw. https://github.com/lllyasviel/stable-diffusion-webui-forge/commit/b835f24a4d9732c936a3d41d62adfae46bd82441
>>30331 >neggles tranny is hanging out here grim
>I merged bait.
Stop posting the same prompt with some minor variation, dumb jeet. This isn't Civitai.
(2.22 MB 1920x1536 20250424131557-2348352433.png)

(2.25 MB 1920x1536 20250424131914-2348352428.png)

>>30332 >>30333 Sorry, GitHub is shit and doesn't show who actually merged but it was me instead. Check here. https://github.com/lllyasviel/stable-diffusion-webui-forge/pull/2151#event-15913934952
>>30334 Go back to treechan retard
>>30337 >it's okay when we do it
>>30331 Ah, nevermind, I guess it's a safe option then. Still, Illya's weirdness surrounding the whole ordeal kinda soured me on base Forge.
>>30336 im a complete codelet what's the issue with obfuscated code? high probability of malware? arent most programs closed source anyway?
>>30340 the rest of Forge is open source as is comfy and A11
>>30340 It's just comfytroons having a meltie over lllya being the superior dev
>tfw Voldy never came back to work on A1111 >Zutin probably drafted him to go die in some dumb war Man, he's probably dying in a trench somewhere in Ukraine right now
(2.13 MB 1344x1728 00048-2434620692-00003.png)

(1.91 MB 1344x1728 00031-1234196170-00002.png)

>>30281 Its a little fucked because it seems they have both the older and newer style mixed in but, good enough
>>30265 Thanks, yeah after tinkering around a bit I figured it out somewhat, the biggest thing was artist tags like you said. Gonna try LCM, I was usually on Euler a. As for no negatives, I'm sometimes getting random scribbles that look like Japanese/white borders, also for some reason it loves to make the girl sweaty/wet skin even if there's nothing in the prompt for it as far as I can tell. Did you run into those issues?
>>30226 It's fine, thanks for replying. I assumed it's probably not used much, figured still worth a shot. I could increase cfg, but if it's above 1 it activates neg prompt and makes it take almost twice as long. I'm trying to see how much I can reduce the generation time while still getting decent results. For now I'll just not use regional prompter, just thought maybe someone knew a trick.
>my pillar hasn't posted once >my pillar was a bot all along bros...
>>30347 Can you drop that stupid meme
>>30349 children all look the same to me
(4.52 MB 1792x2304 ComfyUI_00004_.png)

>>30345 >Did you run into those issues? Yes, the sweat doesn't bother me, I often prompt for it explicitly anyway. The scribbles also don't bother me, I find it amusing. I've gotten very legible hiragana and semi-legible kanji before but it's generally gibberish. I guess if that stuff really bothers you you might need the negative prompt, in which case I'd set CFG to 1.5 maybe.
Astra found a bug in the wip V7 and is fixing it. Supposedly, that's going to make all of the gens look much better and not like sd1.5 anymore. I can't post preview cuz it's a pony
>>30352 Fair enough, wasn't sure if it was a me issue. The sweat doesn't bother me for lewd pics, but it can be awkward for sfw pics. But yeah, not that big of a deal overall I guess.
>>30343 voldy was always scum, you could tell it from how he presented himself
>>30353 >someone said a thing
>>30353 You can post ponies here, probably.
>>30357 no, they're like an extra bad version of furry that said I don't know if there is any moderation here at all
>>30358 Box it then.
>>30353 >>30358 If it's boxed/spoilered and it's for a model preview then it's probably fine. If you genned 5 ponies and decided everyone else likes them as much as you do then you're likely going to get banished. Context is king.
>>30309 I'm honestly kicking myself for having not uploaded a lot of shit I wanted part of the dataset onto Gelbooru (which they surprisingly included for V4 despite its worse quality control than Dan). Didn't think they would do a retrain this soon. Gonna have to get on that shit sooner than later now.
>>30363 better start now for v5
>>30362 should be ankh on middle image baka
>>30365 It's so over ....
(2.67 MB 1536x1920 20250424172524-3980207795.png)

(1.91 MB 1920x1536 20250424224109-3480238259.png)

i downloaded that lora now i too can make generically appealing gens and get twitter bucks extra nipples and weird stomachs are normal for chimera
>>30368 >i downloaded that lora which one?
>>30369 it's nyalia
i realized a critical flaw in my twitter bucks plan >>30369 this one >>30281 >>30370 There's other artist names in there too. Even with the lora at 0.5 the Nyalia is strong though, yeah.
any new models out? are you all on that 102d one?
>>30372 re3mix is looking good so far
>>30372 using my personal shitmix
So, thoughts on the new illustrious models? 2.0 and Lumina 2B
i have no ideas what to gen and it feels like im forcing myself lately
>>30372 why do you need a new model
My way-to-go cloud GPU died. It was a good 1.5 years of free gens Go on without me.
>>30377 102d is boring
>>30379 then don't use it? lol
>>30380 what else is there to use
>>30381 every other slop mix out there
>>30381 boobai
>>30381 vpred 1.0 with loras
>>30381 Pony with autism102d
is this shit new on civit?
>>30387 >posting to civitai did you strip the exif data or something?
>>30388 not me posting, just saw it on some lora example pic which makes it weirder.
>>30387 >>30207 >Require all mature content to include generation metadata It's early days so probably still some kinks in the system. Seems like you can just add a prompt+seed manually even if they're fake.
>>30390 Cryptobros warned us about this but we didn't listen.
>>30390 You will always be able to add whatever metadata you want, though, you just need to edit it. You can also always inpaint, even with 0.9 denoise, with a mask that is 1px wide at the corner of the image. Metadata is obviously not a real solution.
>>30331 >>30336 tldr about why there is so much whining about this commit breaking something?
>>30393 We taked about it months ago. illyasvel took some code from comfyUI, and instead of giving credit tried to hide the fact by obsfucating the code, which is a big no-no in open source. So people began to wonder what his reasons were (other than pride), such as distributing malware.
>>30394 The weird thing is that no one would even bat an eye if it said "code copied from comfy, not my fault it's shit" or something equally passive aggressive. Instead, he takes the insecure approach and tried to hide it.
Why are there so many comfytroons here?
>>30396 i despise cumrag and and his pathetic ui but in that particular situation i really can't find an angle where illyasviel is right, it really was very weird behavior from him
>>30397 The chinaman stole or copied something without giving credit, so what? Next you're gonna tell me the sky is blue
>>30394 >>30395 okay, but it doesn't explain why it broke some sampler functionality for someone https://github.com/lllyasviel/stable-diffusion-webui-forge/pull/2151#issuecomment-2590654866 there was some obfuscated code related to that?
>>30399 The samplers code is exactly the code that was copied from comfy (again, no one is against copying that code, but Ilya can't admit to it for some reason). He added obfuscated code for those samplers and never added any human readable code, which is a problem in many ways. People deleted the things that don't have actual code from the repo.
>>30398 The core of the issue is putting obsfucated code into Forge. Where it came from or what it does is secondary.
>8chan.se got blocked due to request from kraut government its all so tiresome. at least .moe is working
>>30344 Mostly just gens from catbox collections, was not filtered hard and long enough >>30349 No, don't remember where it originates from, probably should merge into one >>30371 lmao
Are we back to not posting images again?
>>30405 I'm stuck with just civitai genning until mid sunday. Tough getting anything of decent quality with limited rolls, and no inpainting.
>>30406 I was once stuck with that too, after my old GPU finally bit the dust after ~8 years, but before my new one arrived. It sucks, friend. Good luck.
(4.01 MB 2432x1664 00235-973851681.png)

>>30405 I went into another yuri/futa rampage the other day, while you guys seem cool with it I don't want to spam it
>>30408 are you posting yuri / futa anywhere? I can enjoy some yuri, too.
>>30405 Blonde gura?
Futa is gay
>>30411 yeah but futa on female is slightly less so
>>30408 I want to see
>>30408 Yuri is fine here, and there's a /d/ thread on this board, although it's pretty much dead Also surprised you prompt futa
>>30410 Nah, just your run of the mill regular shark girl
>>30414 actually tried talking to someone there only to realize the last post was made a month ago
(3.04 MB 1664x2304 00090-3698028428.png)

(3.07 MB 2688x1536 00243-60395768.png)

(4.07 MB 1536x2688 00178-975224577.png)

(3.83 MB 2432x1664 00325-2564615827.png)

>>30409 anywhere aside this thread/site? not really >>30413 well a few can't hurt >>30414 even while prompting futa I remain boring and vanilla
>spicpillar is a fag it's over
>>30419 oh no :( I was hoping 4shit would stay dead for at least one more week.
>>30419 thread made by that shitty grifter.
How long until it goes down again, my bet is 2 or 3 days
Any ideas why this is only producing noise on my end? https://files.catbox.moe/ldtlu2.toml
>>30419 there's zero reason to go back to 4chin should just stay here
>>30425 >Guess I'll stay on reForge for as long as possible I just don't understand why would you switch from it unless there's some sort of crucial lora type released or a new super model that is only compatible with a new UI. You can get most of the functions from extensions, and one of the anons said he'd implement crucial shit to reforge anyway. You don't need to updoot every week.
>>30424 I know, I'll stay here as well, they didn't change their staff or rules, didn't change anything at all, same old boring circus
>>30424 this, thread's been up for an hour and it already looks like an awful place
>>30426 That's true my bro. I won't worry about updooting then. If there's ever some lora I need to use with regional prompter then I'll temporarily switch to old a1111.
>>30425 What do you use it for btw? I don't see much itt that would require regional prompting.
>>30405 Keep posting this style, I need more pics for VT.
>>30425 How do you get regional prompter to work on reforge?
>>30430 Yuri or /e/, especially with 3 characters, although tbf Noob-based models are already fairly good at separating characters' features without the need for regional prompting.
>>30432 You have to update regionalprompter and then it'll work on reforge, it got updates a few weeks ago.
>>30423 Net and conv alphas are higher than dims, which makes no mathematical sense. Alpha can only be equal to dim or lower. Also, not the cause if your issue but I don't recommend using random crop since the captions are not "cropped" and may describe something that is not visible in the cropped image.
>>30424 Let the tourists go back, idiot.
(1.66 MB 1248x1824 catbox_hh9clk.png)

(1.89 MB 1248x1824 catbox_k4i1fw.png)

(1.61 MB 1248x1824 catbox_za4brv.png)

(1.64 MB 1248x1824 catbox_uzs0bj.png)

>>30417 So you just jumping across the boards? I'll just answer you here. All those boxes have one big flaw, AYS is compeltely incompatible with vpred sigmas schedule, never use it, don't know/care what that sampler is, but simple euler will still work just fine https://files.catbox.moe/1be2t0.png https://files.catbox.moe/1p04f0.png https://files.catbox.moe/pt5coc.png Loras you asked, baked with the usual config, go test them: https://upload.lolispace.moe/0b1d8c16b8b6
>>30419 8chan have been my second home for a year, so nothing changes for me. i'm kinda happy 4chan hdg is back so all the schizos can go back.
>>30424 >So you just jumping across the boards? Yes >All those boxes have one big flaw, AYS is compeltely incompatible with vpred sigmas schedule, never use it Interesting, I haven't had any issues with that scheduler on 102d or r3mix >Loras you asked, baked with the usual config, go test them Thank you very much, I'll do it
>>30435 odd, the config is straight from the machina fork ezscrips repo.
>>30439 >Interesting, I haven't had any issues with that scheduler on 102d or r3mix Just means that it's not true vprediction, which shouldn't start with 14 sigma. And that value of 14 is hardcoded by nvidia, it's a little changed eps scheduler
>>30435 dims weren't the problem, still produces noise with switched values. Maybe something borked with my install
>>30428 Really impressive. It's like that shithole is rotten to the core.
>>30441 >Just means that it's not true vprediction, which shouldn't start with 14 sigma. So the shitmixes are so shit they make AYS kind of work? Funny, really funny Btw, does any of the loras have any activation tags?
>>30441 >means that it's not true vprediction What does this mean practically? Those models can generate true #000 darks
>>30444 No, there is no activation tags. Kinda, I guess these models just merged with something, that have more info about what to do on a lower sigmas, which makes them more optimized, I haven't tested them, I just don't like mixes, it's harder to "configure" model for yourself with loras with any kind of mix, basic huge finetunes are better at that >>30445 It's not a test you should trust, eps1.0-1.1 can too. The true effect should be by not filling anything at the canvas with far shot of a character with a simple bg, night scenes without damn yellow unprompted lantern, etc. Practically, you should just don't care about the math and all that shit as long as it works for you and doesn't make some bullshit like this https://files.catbox.moe/3inyg6.png
>>30445 nta but one trick I found was doing img2img at 1.0 denoise, anything other than true vpred+zsnr will still be affected by the underlying image to some degree. In more practical terms, it's not just about being able to do black or white backgrounds on demand, that's a meme test that even old Reweik v12 could pass. It's also being able to handle many similar shades in one image. A girl in shadow in a dark room with no light sources. An albino girl naked in the snow. A tanned farm girl in a wheat field at sunset. etc. Eps-pred will always add something off to "balance" the image's overall color palette, the "local lamp" that NAIfags make fun of.
>>30446 I could never get V3 to do actually dark images, I had to wrangle it pretty hard to not generate a lamp but the subject was always lit one way or another. Skill issue? Maybe it's not a proper vpred model?
>>30448 It could be artist mix issue. I was experiencing simillar problems with some styles, while the others worked out fine with noob
>>30448 Also I can't say for sure if the proposed vpred implementation is fully correct, even on nai, because shit like this happens even on v3, some prompts with particular seeds just completely burned, while the others will look fine
>>30450 >because shit like this happens even on v3, some prompts with particular seeds just completely burned, while the others will look fine That was discovered pretty much immediately, sometimes things will just come out red for whatever reason.
>>30435 >which makes no mathematical sense. Alpha can only be equal to dim or lower. Where did you learn this dumb shit? Dim/Alpha is just a simple scalar on the weights. It's like setting Lora strength to >1.0 when you set Alpha > Dim (or setting strength <1.0 when Alpha < Dim)
>>30453 Yeah I was wrong, but it seemed a bit awkward to admit after he already found out the issue is elsewhere. Could still cause a NaN in theory if the lora ends up learning too fast because if it, and I can't imagine what effect it has on the original usecase of avoiding rounding errors.
>>30452 >It's a skill issue. I've imported the metadata and it's pretty hit or miss. If you add a character tag it stops working altogether. It's *a lot* better on V4 (far from perfect though) but for some reason it'll almost always give you a shot from behind by default.
>>30455 I guess because you don't describe her face
>>30456 Yeah, a simple looking at viewer fixes it most of the time (though it might still give you a "looking back" kind of shot lol) Just funny that V4 loves defaulting to feet and ass (unfathomably based)
>>29491 metadata, p-please?!
>>30457 v4 is a 4 sided die with each side being: >extreme ass close-up >extreme feet close-up >temporary cataracts >your gen
>>30459 This is 100% accurate. I still don't get why just adding "portrait" or "close-up" quintuples the chance of #3 happening.
>>30459 I laughed. It's funny because it's true.
>>30417 extremely based girlmaxxer
When using number-based emphasis on V4 2 and above almost act like a negative value, what the fuck?
>>30463 "fixed"
>>30464 I hope V4.5 will actually fix this mess of an architecture.
>>30463 "Emphasis", just like "negative prompt", is a big misuse and misunderstanding of the underlying process. The reason it breaks is because you are scaling the vector beyond feature space until it points to nowhere and therefore loses all meaning.
>>30466 TIL. Also negative prompts never ever fucking work right, they might as well not be a feature.
>>30466 >>30467 are you sure? putting "worst quality" in negs improves anatomy (no extra limbs) for me
>>30468 >gloves and fingerless gloves in negs >still get gloves You tell me.
>>30467 Negative prompting works to some extent because CFG will actually try to guide the generation process "away" from the features present in there. But it's not a magic box that tells the model "don't generate things I list here". I use if very rarely to eliminate certain things that cannot be prompted for properly. "monochrome" in negs almost always works as you would expect, and some other tokens, too.
>>30470 >But it's not a magic box that tells the model "don't generate things I list here" ...why not though? Why doesn't it work anywhere near as effectively as the positive prompt box?
>>30471 Put your negative prompt in the positive at CFG 1. Generate. Then ask yourself "what would an 'anti-image' to this look like?"
>>30471 because the negative prompt doesn't do that, it generates a second image that it uses to guide the first image away from it. if you want to exclude concepts entirely use negpip but that also has unpredictable side effects.
>>30472 oh. >>30473 >it generates a second image that it uses to guide the first image away That's retarded. Why does any of this work the way it does?
Any artists with shading (and details) as good and complex as this? https://files.catbox.moe/q4x5it.jpg https://danbooru.donmai.us/posts?tags=arenbee&z=1 I don't think VT is gonna cut it...
>>30476 momisan
>>30477 Good suggestion. Unfortunately while the torsos look pretty good the faces are really scuffed on V4, need to go all the way down to 0.5 and at that point the default V4 style(s) start to leak in. I could use VT but unfortunately it refuses to do dark images (even if the input image is dark)
>>30376 Idk if you're still here, but do some magical girl cosplay. Baiser is pure sexo.
>>30475 >Rewards >A 4chan Pass valid for one year ($20 USD value, subject to Terms of Use). waoh
>>30480 it's so fucking embarrassing
>>30474 it's just a computer, it can't reason. all it has are patterns of shape and color
>>30475 kek also we back?
>>30483 yes, fedchan /hdg/ is back being complete fucking shithole filled with subhuman retards
>>30484 i'm gonna keep posting here sometimes i think, its nice
>>30484 so much for hope it'd improve after a break somehow its even worse now
>>30486 they do post images tho, which is kinda nice
fuck I always click Archive instead of Refresh because that's where the button used to be
i'm very glad that one dude is still posting a lot of slop there but...
>>30474 anon if you're gonna bitch about it you can fuck off back to halfchan or take a moment and realize it's better learning how it actually works than thinking it's pure, flawless magic like how every other AI product is marketed. it works like that because SD never originally had "negative" prompting. the negative prompt feature we have in the webui was cooked up by A1111. if you want it to work, this is something you're best learning and experimenting with. take the time to understand it. try different models and extensions. learn by doing, instead of getting others to tell you how to do it.
>>30491 I think you should cool off, pal.
>>30492 sorry for being so (((aggressive))), i haven't had my soymilk latte this morning. i never feel very tolerant until i've had at least two
>>30491 lila box?
(1.81 MB 1200x1600 catbox_vqc7si.png)

Sometimes I want to fap with simple sakura.
halfchan comes back up and thread returns to a snails pace. at least we have reached baseline levels of autism now
>>30496 It was pretty obvious most people will just come back there. No matter how objectively shit 4cuck is nowadays, how retarded the rules and jannies are, everyone wants to post to an active thread. Honestly I'd prefer halfchan to die for good even if it meant that some shitposters would move to altchans.
>>30497 I feel like those same shitposters wouldn't get as much leeway here. Dude can literally post dall-e scat on /h/ with seemingly no consequence, and it'll stay up for hours.
>>30496 It came back too quickly and it's vey sad. Needed to be dead for at least a few more weeks.
>>30496 a-anon, your VAE?
(1.58 MB 1536x1536 20250428070945-3242027054.png)

(1.66 MB 1536x1536 20250428065607-3240299097.png)

>>30501 box?
>>30503 homu version of that pls
(1.81 MB 1536x1536 20250428103453-2330487957.png)

i wonder how long it'll take to push out v4.5 realistically it should be another month at the very least, no?
>>30507 wait what, I thought v4.5 was already out?
>>30508 a preview/early epoch is probably available to a few big xitter/pixiv accounts but definitely not to the public
>>30507 two months until curated
fingers crossed i guess
>>30511 Before they said v4 would be the new v3. Still not convinced it's possible for one person to bake two good models, it's always just a fluke and their success-boosted ego stops them from seeing it as such.
>>30512 >Still not convinced it's possible for one person to bake two good models, it's always just a fluke and their success-boosted ego stops them from seeing it as such. i think it's a little early for that kind of snark, only time will tell
>>30513 That is how it's been so far, based64, easyfluff, pony, 4th tail, illustrious, naiv3 they all tried to ride their wave of success and bake something even better, only to fail miserably.
>>30514 naiv1 was pretty good for its time too though, it was a big deal. so they have v1 and v3 as their successes. but so far right now its looking like they're just running damage control for v4. im personally happy with local now though so i'm not really invested in this
(813.37 KB 1200x1700 homu.jpg)

>>30518 https://www.pixiv.net/en/users/7222503 Anyone able to recognize this style? 1st picrel as example. I know it has both kumiko shiba and kakure eria but I cannot pinpoint the remaining one. Reminds me of moisture (chichi) but it's even softer/muted than that. I've definitely seen it before and my memory is failing me. 2nd pic unrelated and so I'm not just only requesting something. :^)
>>30519 Fuck, didn't mean to reply, sorry; have another gen as an apology.
>>30500 i don't think it was a vae issue, i think it was a CFG extension i'm using. i suspected at first it was the prompt i was using like 'morning' and 'sunlight.' guess i was too wasted to really notice
>>30520 sloppy sweaty anal with nitori...
>>30519 eyes are def asanagi
>>30519 maybe tomiokasena
damn i missed a lot
(441.92 KB 512x768 250429_201659_963_5700.png)

(1.86 MB 512x768 250429_200424_508_7537_37.mp4)

(187.55 KB 544x704 250429_153900_327_2913.png)

(803.58 KB 544x704 250429_153900_327_2913_19.mp4)

framepack is kind of fun 1:40 per second of video is rough for experimenting though
>>30527 Is a speed bump really a good reason to get ripped off that hard?
>>30530 >No but the ability to run model 2x ~ 4x the size of SDXL might be Let's be realistic here anon, we're not escaping SDXL for at least another two years. Might be decent for video stuff but all the img2video/text2video results I've seen don't impress me, they all look like fiverr L2D jobs.
(2.60 MB 1536x1920 20250430065235-2674079042.png)

>>30519 I messaged him and tuned out after reading the first line of his reply. Oh well. >The complete combination of artistic styles is a secret formula I can’t disclose, but I can share some of the styles I use for your reference: besides the ones you already know—Kumiko Shiba and Kakure Eria—there are also Asanagi, Mamimi, Happoubi Jin, Hews, Lasterk, Afrobull, and more... Additionally, the LoRA models I trained myself (based on works by lesser-known artists) play a significant role.
(3.69 MB 1664x2304 00130-1153319464.png)

onee-shota is a very based tag ngl, it gave some gen ideas for later
>>30534 rubbin legs with the bros while you nut
been out for a couple months, what's the trendy flavor of illust/nai to run now?
>>30536 No one really use any illust based models at this point, we use noobai vpred (shitmixes or base model with some loras)
>>30537 thanks, I'll just redownload wai and r3 shitmixes for now since I deleted all my models
The realism tag can be so dangerous It'll either improve the image with more realistic proportions and look or everything starts glowing and sirens start blaring Aren't these models trained strictly on anime? How come the result can sometimes end up like THAT?
>>30539 >afraid of the realism tag you would've never made it in the no-no model days
>>30540 I'm exaggerating for comedic effect but it really does feel sometimes that the picture goes from oh this is good to FBI AGENT I AM SORRY IT WAS NOT MY INTENTION
>>30542 god, deep penetration is so good but can be such a pain to prompt
>>30543 (balls deep, deep penetration) seem to work 99% of the time for me.
>>30543 there are sliders/loras for it if your mix is unstable with the appropriate tags helps to not prompt penis out and neg penis out on another topic >v legs >wide spread legs wish I knew theses tags holy shit legs together keep sneaking in in some mixes
does anyone have this guy's as109 char loras? https://civitai.com/user/sg5d civitai is cucked as shit
Using FramePack-Loop on an image output by Stable diffusion
>>30548 interesting, does it work well with sex/hentai?
>>30548 That is awesome. How long to gen? Need to learn to do this next.
>>30551 >i'm a beta tester! >absolute promptlet why does this keep happening?
>>30550 for me it's a little under two minutes of gen time per second of video but i wasn't getting good sex results so i wonder how he's prompting it >>30552 average user is a promptlet too
>>30548 fucking based.
>>30555 >Model: test1 Is this model published somewhere? None of my gens with the same artist mix looks as good as yours
(59.51 KB 1442x598 error.png)

I'm trying to train a lora using the colab method with easyscripts but whenever I press the training button it loads for a few minutes then gives me this ^C thing that stops it. Did i do something wrong setting everything up? I thought I followed the guide pretty well.
>>30559 Is this free tier? I don't think Google colab free tier has enough RAM to train SDXL
>>30561 oh really? damn. I went through all of that to set it up too.
are we still posting here, please tell me it's active because holy shit halfchan hdg is insufferable
>>30563 This place has existed long before 4chan's downtime and would have kept chugging along regardless. It's never been super active but that's been one of its boons imo.
(3.33 MB 1664x2304 00214-2855613731.png)

>>30563 some people still post here, now if you want those sweet (You)s and/or interactions, that's another story
>>30558 It's a noob v1/v29 frankenmix so nothing special. More specifically a cyberfix/itercomp mix plus a v29 mix https://mega.nz/file/aYkHySLS#KYsLLanu3Ycnf3iWyUU2sb7-Rf81-Xb_-D5XIx1NzlU Just make sure you've enabled APG
>>30566 very nice
(2.72 MB 1536x1920 20250504155534-2239314408.png)

>>30569 Box please?
Suggest some artists with realistic/detailed skin, they're my bread and butter but I'm drawing a complete blank for some reason.
>>30571 Do you mean like hizuki akira or padoruu or more realistic like neoartcore
>>30572 Akira is alright but definitely more towards the other two. The ideal would be neoartcore but without the more realistic faces. eu03 is also good but most of his pieces are basically unfinished aside from the bodies (I'm hoarding VT material)
>>30566 box?
(12.25 MB 3696x10970 file.webp)

(10.31 MB 3696x10970 file.webp)

>>30567 >APG Turns out this makes all the difference, not the model Sadly APG in my limited testing almost always gens degraded hands
Any poorfags out there figure out how to get FramePack to work on 20-series card? Inspired by >>30548 but don't have the means of upgrading any hardware right now
>>30576 did not bother with it. it uses like 48 gb of ram (or swap memory). probably will be very slow.
>>30577 tried xformers and it errors out saying the card is too old, fml
>>30563 halfchan hdg is same as it was before the hack, complete garbage. I really wish 4chan didn't come back at all tbh, altchans are gonna be semi-dead again
>>30575 Are you using default APG settings? It might help to tweak them a bit even from my settings since they can vary by model/sampling/scheduler/etc.
(8.29 MB 8192x7168 file.jpg)

AI styles trained on AI styles trained on AI styles
>>30567 nta, but thanks. nice mix
>>30580 Copied most of your settings verbatim and now it's much better, thanks Guess my CFG was too low
new thread maybe? so that hdg is at top
>>30584 i don't see why that would matter
>>30586 dont really wanna see the yaoi pics and cant hide the threads as I am mostly posting here from browsers that wipe cookies on exit. not a big deal of course
>>30587 use one of the adblockers with a content zapper, this board is slow enough that'll last you for a good long while
>>30588 a new thread wouldn't hurt either. we'll call it /hgg/.
>>30589 >we'll call it /hgg/. why?
(3.60 MB 1536x2560 00164-2243833022.png)

>>30591 Oh, is bestiality ok here?
(2.14 MB 1248x1824 catbox_sdpow0.png)

>>30592 Dunno, that was a test prompt for this.
>>30585 Thanks anon, very cute
>>30591 damn, those are both peak
Definitely better at mixing, will have to see when they add the nsfw stuff. Still has some inconsistencies but, nothing extreme like it was before.
I haven't measured exactly how long it takes to generate one video, but it takes about 10 minutes to generate 2 seconds, and it's very tedious to try again and again until you get the right answer...
>>30593 nice image but why censored
>>30596 Either mixing still doesn't work or none of the artists I'm trying made it into curated.
>>30600 Works fine with mine.
>>30601 Do traditional media/(medium)-related tags work correctly on your end? Depending on the character it either spits out something that's very obviously unrelated to the tags OR it spits out monochrome millipen-esque stuff (that's very nice but again not what I'm asking) without failure.
>>30602 I don't use traditional media much, but what little I tried seemed to do okay. Maybe quality tags or UC are getting in the way? Or prompt guidance? When it goes monochrome, I just hit it with a -1::monochrome:: that usually makes it behave for me.
>>30603 > -1::monochrome:: Nope, didn't help. I even disabled the UC preset and auto-quality tags.
>>30604 It's also nice to see that they once again didn't fucking bother fixing DPM++ SDE at reasonable guidance values
So what's the best model for cunny these days?
>>30584 People usually bake at around this posting count so should be fine
>>30606 local? 102d_final
>>30600 Smaller artists aren't in I think, or exclusively nsfw artists arent either you can test them just by themselves
>>30609 I tested fizrot and he's not in unlike the old curated/curated preview
>>30610 Are you sure about that?
>>30611 Positive. The tag does *something* but that's obviously not fizrot(art). Try it on V4F/C or even V3.
>>30613 Only the third one is fizrot. For some reason V4 sees "fiz" and wants to spit out fizintine instead. I just neg it when it happens more than a few times in a row.
>>30612 V4F and V3 probably look more fizrot, but I'd say V4.5C does a much better job than V4C.
>>30615 V4C is overfit but V4.5C is extremely underfit, even more so than V3.
V4.5C >>30616 If V4C was overfit, it would look a lot like some specific fizrot image, but it just looks like a completely different artist. V4.5C is a lot closer to the correct look. Even the third one there looks less like him than any of these. They're just all kind of blue and purple, but at least they have some glowy colors, compared to none on V4C.
>>30618 >V4.5C is a lot closer to the correct look. Glasses. Now.
Oh, there are only 12 sfw images, that explains the difference in the curated models. >>30619 The only resemblence that third image on V4C has is that the eye is kind of pointy and it's painterly. The sense of colors and contrast is completely different.
>>30618 curated literally looks nothing like fizrot
>>30621 V4.5C has intense colors and it's painterly. That's more than can be said for V4C. I agree it's not actually close to his style.
>>30620 >only 12 sfw images The earliest version of v4c already got it working. Either they massively fucked up the dataset that time or they fucked up this time. >colors and contrast are completely different Congratulations on wasting 3 minutes of my time but we're done here.
>>30613 >The earliest version of v4c already got it working. How can you call any of these "working"?
>>30610 he wasnt in the old curated he's like 90% nsfw art so he wont show up until full, if you look on danbooru all but like 4 or 5 images are gold locked
>>30625 He was. Like I've said, V4 likes to spit out fizintine instead of fizrot pretty frequently. Also they scrape from gelbooru apparently.
FramePack is fun, but I'm not sure even a single 4090 is enough to do this in a speedy manner
eat my shit captcha
(731.92 KB 720x1280 WANI2V_727881284636245_00001.mp4)

>>30627 i switched to wan 720. fucking nine minutes to gen two seconds of video but it's got about a 25% hit rate.
>>30521 >>30557 Box please? >>30567 >Just make sure you've enabled APG Thanks for bringing this to my attention.
>>30548 >>30597 >>30598 >>30627 >>30629 >>30630 framepackfags, are you genning using the web demo, framepack studio, comfyui, or something else? do the base models generally work with lewds or do you usually need loras?
>>30632 i'm using illya's tool for hunyuan and the comfy workflow from the rentry in the /g/lmg/ op for wan pin-up stuff works fine with a base model but my few attempts at penetration turned out pretty poor so i'd think it needs a lora
>>30627 jank, but cute!
>>30632 Base Illya webui, but as other anon pointed out the hunyuan model is quite limited, you basically have to trick it into doing anything nsfw
>>30635 unrelated but do you have that takorin lora?
tomoe hotaru.
>>30639 that's not too bad but the juices are zwooping in and out
I want to throw those twin train brats into this prompt but it seems like I need a LoRA to gen them.
>>30642 you mean hikari \(blue archive\), nozomi \(blue archive\) ? I have no troubles prompting for them on noob base v10. iirc you're using noob v0.75s, maybe its because of that?
>>30642 Try regional prompt >>30525
>>30643 Wait wtf it's working now and I used the same tags as when I tried it the last time. >>30644 I will if I need it
(2.30 MB 1824x1248 catbox_d2kjou.png)

(2.37 MB 1824x1248 catbox_y9q47p.png)

>>30646 Yeah you are correct and I should start using regional prompter again, as I want to "remake" some of my old 2 choco maid images as well.
>>30647 welp Hikari didn't get pointy ears in the first image.
How good is regional prompter supposed to be? It didn't prevent the leak of neither expressions nor costumes in my tests
>>30647 anon... don't forget the halos, please
>>30650 And the tail, I have already tweaked the prompt.
I couldn't get dynamic prompts to work on Forge anymore, so I switched to comfy. Does anyone here use the wildcard extension on comfy? I'm trying to get it comfy for me, is there a node or whatever to extract the prompt from the finished picture? I also want it to save the actual prompt, not my brackets and wildcard file names. I can't find the setting for that anywhere.
>>30649 >akari Good taste.
It's hard to get a better one in generation.
I wanted 2girls to lick up from the base of the penis but it didn't work. Sorry to keep posting similar items.
(1.45 MB 1152x2016 file.webp)

(1.14 MB 1344x1728 file.webp)

(1.39 MB 1344x1728 file.webp)

The best part of deep penetration for me is when it causes the girl to puke
did stealthpng stop working for anyone else?
>1128 posts One of you should really make a new thread
What did she mean by this?
>>30659 its hard to prompt for deep pen for some reason
(2.45 MB 1536x1920 20250513152228-3447237753.png)

(2.53 MB 1536x1920 20250513152446-3447237751.png)

(2.31 MB 1536x1920 20250513152559-3447237751.png)

(2.49 MB 1536x1920 20250513152901-3447237751.png)

(2.94 MB 1536x1920 20250513153401-3447237757.png)

i forgor
(11.89 MB 3344x10932 file.webp)

>>30668 "balls deep, deep penetration" works for me more often than not
>>30672 yeah i did some testing and i think its artist dependent more than anything
>>30671 CUTE
>>30670 very cute
(2.73 MB 1536x1920 20250514115014-2404817876.png)

(2.35 MB 1536x1920 20250514180930-3447237763.png)

>>30679 Interesting way to use regional prompter >bird's-eye view, high-angle view These are e621 tags right? Do they still work in recent merges?
>>30680 Unless recent merges have somehow done extensive training excluding e621 dataset then I don't see how they would work any differently than noob
(1.69 MB 1248x1824 catbox_tnu0ba.png)

FRESH BREAD >>30685 >>30685 >>30685


Forms
Delete
Report
Quick Reply