Disappointment edition
Previous Thread: >>
>>8653057>LOCAL UIreForge: https://github.com/Panchovix/stable-diffusion-webui-reForge
Comfy: https://github.com/comfyanonymous/ComfyUI
>RESOURCESWiki: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki | https://comfyanonymous.github.io/ComfyUI_examples
Training: https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainer
Tags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag
ControlNet: https://rentry.org/dummycontrolnet | https://civitai.com/models/136070
IOPaint (LamaCleaner): https://www.iopaint.com/install
Upscalers: https://openmodeldb.info
Booru: https://aibooru.online
4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb
Illustrious-related: https://rentry.org/illustrious_loras_n_stuff
Useful Nodes/Extensions: https://rentry.org/8csaevw5
OP Template/Logo: https://rentry.org/hgg-op/edit | https://files.catbox.moe/om5a99.png
>>8660641No idea what happened to the last thread, more than likely some retard doing the usual stuff to derail the IA threads but whatever, I already had the bake so there you go
>>8660641it's all ayasloppa
Finally, I was starting to get worried this was the end for us. I've started to get attached you you lora baking anons and would be sad to lose you all.
>>8660643main thread got spammed so hard mods nuked the whole thing
probably took this one along too since it's a pointless duplicate
>some idiot needs to make a dozen new threads to push one to page eleven
>just make a new one and nullify his efforts
Can somebody explain this?
>>8660641ayakon why would you do this
Slandering our Lord Ayakon like that is unforgivable.
>>8660648dude you're in the wrong thread, this is /hgg/
>>8660649he's jealous of cathag
>>8660641shitters are getting boring tbqh
>>8660641Lack of beatings from fatherly figure.
>>8660662isn't that a requirement for being on /h/
>>8660665why do you think the average /h/ shitposter is afraid of hags
>>8660667Explain how that is a hag and not a JK
>>8660661Box? Really like the style
>>8660670It's a lora I trained on https://danbooru.donmai.us/posts?tags=ruhee&z=5, but I'm still not 100% sure I'm done rebaking since I still feel there are issues.
>>8660674was gonna say it looks like ai slop with all that shine but that also sorta seems to be the artists style lmao
>>8660677Korean artists tend to do that.
>>8660644hehehe nyo! suffer!
bake hdg or i will have to shit up this thread instead
>>8660681you really think so?
>>8660682kurumuz's orders were very clear, rad the email he sent again
>>8660685dont you have some threads to post on the board to keep slidin'
where's my /hdg/
I don't like /hgg/
Any spicy request that I can catbox? I need to take my new settings out for a spin.
>>86606901girl,fern,spread anus,detailed anus,puffy anus,big anus,anal hair,y anus
>>8660681And nothing of value was lost
>>86606901girl, kanna kamui, huge ass, anal, spread anus, puffy anus, loli, excessive cum, huge penis,
>>8660690male focus,otoko no ko, xingqiu \(genshin impact\),small penis,anal,ejaculating while penetrated,pov,paw pose,fake animal ears,cat ears,neck bell,jingle bell,animal collar,paw gloves,white thighhighs,skindentation,thick thighs,tongue,narrowed eyes,anal orgasm,looking at viewer,large insertion,^^^,stomach bulge,heart-shaped pupils,detailed anus,puffy nipples
>>8660695https://files.catbox.moe/lfge81.png
>>8660690what are your New Settings
immiscible_noise or edm2???
>>8660697https://files.catbox.moe/b9utqm.png
>gen on snakeoil fine tuned model for one artist
>use equally snakeoiled lora of that artist in addition
>have lost all recognition of what that artist actually looks like
>endlessly trapped training and piling up snake oil to chase the essence of some imaginary style.
>still coom
>>8660643>to derail the IA threads>IA threads>IAYou rang?
>>8660709nice eye melt
takes serious effort to break them that badly
Inteligencia artificial de hecho
Chat deberia postear unicamente en español de ahora en adelante? sacaria unas buenas curas
>>8660719and a buenos dias to you too
>>8660721juan gorbillion loras
>>8660712Whoever made that batch did not seem to care tho. I still take it for the alleyways and some nice ideas here and there.
>>8660715Yeah. As if her name was already a hard search, people spamming IA as "Inteligencia Artificial" and related stuff to it make things even more annoying.
>>>/vg/531190984
It has been confirmed that L-A-X, the creator of NoobAI, is a NovelAI shill.
>linking a crosspost of a crosspost
ok
>>8660744mental illness knows no bounds
>>8660691https://files.catbox.moe/eak4a0.png
pretty bad prompt, even though I fixed it a bit.
>>8660695https://files.catbox.moe/odwk27.png
Really confused prompt, you really need to add in a "after sex, pussy" or something here to make it work.
>>8660697https://files.catbox.moe/szv218.png
Very good prompt, other anons should take notes, turns out well on most seeds.
>>8660702>>8660733See metadata on any catbox.
>>8660741the non-retarded anons were already saying there was no future to neta, so predictable.
>use lora for "stabilization"
>it destabilizes upscales
understood
>>8660761anyone who uses that term is a midwit
>come to local ai-training thread
>ramble about novelai
many such cases
/hdg/ is getting spammed by indians because someone removed the nai ad from the op again
wouldn't they be turks, not indians?
how do i unturk from this indians sub reddit?
>>8660787Just subscribe to NAI.
What's your AI comparison of the same character over the years?
>>8660863whys it gotten worse
meh
md5: 076f683a221125a8596494bcc2a50cb3
🔍
>>8660863>2022>anime>2025>2.988d butiful realistic hentiwhat went wrong
>>86608752022 couldn't do the character
>>8660881give it a lora and it'd BTFO the 2025 slop
>>8660863Why the downgrade?
2022 justs needs a character l*ra btw
>>8660886that's what 2025 is...
>>8660890so whys the style look worse
>>8660863>>>/e/2935795Also, did something happen to archived.moe? The search filter by thread subject doesn't work anymore, and image links are dead.
eyup. classic soul and soulless situation right about now.
anybody in possession of a noobai fate testament lora
>>8660905use archiveofsins instead for /h/
archived.moe is shit
>>8660905Image links have been ded for a while now.
>>8660911>fate testamen...dark illya?
>>8660890TRUUUTH NUKE.
local in 2025 is just 2022 but with loras. local has not advanced at all. in fact, it regressed!
>>8660910post yours cmon
is the whole thread just two guys or did everyone join in 2025
we're in the same limbo currently as we were when SDXL released. many months of people failbaking new architectures. only thing is, the new architectures take much more time, effort, and skill to train effectively.
>>8660932im not at my pc with my old gens on it
>>8660910You just gotta cherrypick better
>>8660931I didn't even used local, I used civitai and added style of some artist
my potato pc cannot generate on local
>>8660933new architectures =/= better architectures
trannyformers suck
>>8660939lumina 2 is genuinely good doe
just need someone who knows how to cook (not happening)
>>8660938civitai is local. tensaaarart is local. SEAart is local. india is local.
>>8660938>potato pc cannot generate on localsaar, there are places more fit for you
india is localized inside anon's mom pussy lmao
>>8660914>>8660915Oh yeah, thanks. I can just switch 4chanX over manually, thought that would have to come in an update.
https://x.com/elonmusk/status/1944710462165447104
local when
>>8660914I use 4chan-tx and it is still broken, what gives?
>>8660951local is decades behind at this point really
Now that the AI general is way past its prime, where's the general for finding a gf?
>>86606901girl,1other,german shepherd,close-up,sasha \(haguhagu\),blue eyes,forehead,blonde hair,loli,flat chest,nude,uncensored,french kiss,facing up,(head back:0.8),from side,sideways glance,narrowed eyes,looking at viewer,smug,smile,open mouth,middle finger,dog,bestiality,saliva,black nails,fake animal ears,dog ears,bracelet,animal collar,dog collar,bone collar,nipples,blurry background,apartment,blush,aroused,sweat,
>>8660952tx is somehow different from just x?
>>8660954Then why does it still make the most anime images?
>>8660951Isn't this the guy who is about to get sent back to Africa?
>>8660960Can't stand him either, but way worse people have stayed. The prez' can't do whatever he wants, it's not a dictatorship yet.
>>8660959>sdxl blursaarsloppa is anime!you will need to provide proofs
>>8660963v3 was the only blurry sdxl model tho
>>8660951what the fuck does this mean? "waow cool 3d gurl"
>>8660968It means "please buy more blue checks, this piece of shit is losing too much money and I thought I could use my donations and control over social media to influence the orange nigger's politics and I want the x (formerly twitter) acquisition to not be a total net loss".
>>8660956https://files.catbox.moe/gvp1dm.png
>>8660956https://files.catbox.moe/dj5vsf.png
Good prompt, although in most gens the dog is super anthro (e621 strikes again).
>>8660971>although in most gens the dog is super anthro (e621 strikes again).Maybe try adding "feral"
>>8660974I like to think than rather than being AI, Musk just hired a thousand Filipinos, put a motion capture rig on them and they are all in some buidling in front of a green screen each with a teleprompter telling them what moves the anime girl should do and then just runs the motion data through some VR Chat thing.
>>8660956>>8660972https://litter.catbox.moe/x1nocwlpoadu6ud2.png
feral is pretty weak as a tag, probably "german shepherd, 1other" that's the problem.
>>8660973what caused this and thats and how did we get here
the cause:
>>>/pol/510380887
>>>/pol/510381153
Sasha-chan fucking dogs on camera fits with her overall character arc to be honest.
>>8660980>noobvpredv10_made_in_abyss_v1did you share this lora already 'nonie?
Does anyone here do requests?
>>8661007Could you generate a lewd of my OC or is it established characters only?
>>8661016Maybe something like this?
>>8661016If you do it, her hair is short.
shit is the other place really dead? No wonder this place is blowing up. Wtf go back you dumb retards.
OCEE
md5: d0bab67c97dd8bf9b17d1258e170bbf3
🔍
*raises paw*
what's "the other place"?
>>8661044Thank you! It's cute but the body is a bit too petite... Is there a way to generate a slender, but curvy woman?
>ComfyUI is a piece of shit geared towards turning your PC into a fully automated slop factory while basic shit like inpainting requires you to jump through hoops to achieve
>Forge/reForge are dead
>ForgeClassic makes your gens look terrible
P-pancho.. onegai.. onegai kusadai
>>8661121what new features do you want on reforge?
>>8661125All the new snakeoil samplers/schedulers.
so glad i subbed to nai lol
>>8661140I dunno. I see weird ones from when the cumfy users dab on us reforge peasants with their schizo spaghetti metadata.
>>8661144we got too cocky local bwos. nai is getting stronger.
>>8661147now link or post their gens using said spaghetti
heh, they still haven't noticed.
holy shit they actually left
What happened the hell happened, what did I lose
>go on pixiv
>search for X fetish
>it's all shitty slop
>filter AI out
>it's still all shitty slop, because they just can't help post their garbage drawings onto the internet as if people care
I fucking hate humans so much.
>>8661232I hate subhumans like you who can't draw more, however.
>>8660970>>8660971Where the fuck did the images go
>>8661236Not everyone has the time to go deep into every hobby that exists. At least I don't post my garbage I do draw onto the internet like a shitter.
>>8661121>while basic shit like inpainting requires you to jump through hoopswhat hoops? the inpainting UI at this point is better than in forge, opening the mask editor expands to the entirety of the window rather than being a tiny embed for ants like with a1111 derivatives
you mad because you need a separate workflow is that it?
>>8661257I'm mad because I FUCKING HATE NOODLE SLOOOP!!!
>>8661257The mask editor in comfy is a clunky piece of shit and you can make the inpaint interface in forge full screen, you ungenuine troon.
>you mad because you need a separate workflow is that it?Yes, it's clunky and retarded and obviously against the entire design filosophy of comfy which would rather fully automate the process with no input from the user after you are done building your personal spaghetti sculpture.
>>8661297>The mask editor in comfy is a clunky piece of shit>open mask editor>paintdone
do i need to spoonfeed you a workflow or something
can someone explain why people decided to use the trainwreck that is Gradio in the first place anyway? Don't tell me that it's just because it's python
fellas, any model better/consistent for creative sex poses/angles than 291h?
>>8661316one anon is cooking a finetune extract
>>8661316>consistent>creative?
>>8661316it's... it's...
sigh
"It's pretty hot out today, isn't it? How about you go take a break from guard duty for a bit and go have a drink, hmm?"
>>8661327are we supposed to hate lax now because he said local isn't close to 4.5
>>8661324You'll never be the wordsmith Sir Solangeanon is.
>>8661342>nowkeeeeeeeekerino
hace the retards migrated from hdg?
>>8661383Did you mean "Have"?
But yes, both threads got nuked at the same time and the other took half an hour to rebake so people moved here.
>>8661342>are we supposed to hate literal paid novelai shills spouting bullshit lies about local?
>At least I keep being a nogen nodraw
what an odd thing to be proud of
>>8660999It's not my lora, some other lora baking anon shared it a while back. I could reupload to litter if you want it.
>>8661383As long as we keep this thread slow enough that it's boring to troll we will be safe.
https://litter.catbox.moe/fd2q7oca56c4ryuu.png
>>8661399yeah, I just fat fingered
whats the difference between /hgg/ and /hdg/??
>>8661450they have different names
>>8661450Even if you're trolling, might be useful to anyone else who's wondering,
/hgg/ was split from /hdg/ due to the excessive amounts of shitposting that takes place in /hdg/. It's become a slower thread that mostly focuses on the technical aspects of models, lora training and workflows, while still sharing gens from time to time. You can post gens here, but don't be surprised if lots of anons mostly theorycraft in this thread.
>>8661414Both threads are already that slow and then some.
I woke up this morning and saw 15 new posts on the /hgg/ counter and 40 at /hdg/.
I used to wake up to 800+ new posts in 2024 and had to skip most of it because catching up was pointless and insane.
>>8661469Shit gets old, people get bored, nothing particularly exciting happening, genning slop is no longer new as it was back then. Even shitposters got bored for the most part.
>>8661469>I used to wake up to 800+ new posts in 2024 and had to skip most of it because catching up was pointless and insane.I mean, that was still /hdg/ a month or so before the split. Just that 30% of those 800+ posts that were anons discussing models, sharing gens, helping out, etc. in 2024 dropped to about 5% before the split. No one even quotes the posts their responding to because it's some sort of anti-troll tactic so you'll have 50 posts of incoherent verbal diarrhea and it just breeds further shitposting.
>>8661414stop lewding shigure or im gonna have to jerk off
>>8661472>No one even quotes the posts their responding to because it's some sort of anti-troll tacticTroll tactic actually. It's like trashtalking someone behind their back but loud enough for them to hear. As opposed to criticizing them directly.
>>8661476this did start as the "anti-troll" tactic (I'm not giving shitposters (You)s!) but yeah eventually it became another way of shitposting too
>>8661476>>8661477Yeah you're both right. Just goes to show that's where the general was headed anyway.
>>8661469>people got tired of all the shitposting and just stopped visiting>shitposters don't care about a thread if it isn't active so they stop propping it up as much>thread is now slow>with it now being slow the real constant schizos are still there and filling the thread with their garbage >people continue to not want to contributeI'm personally at a point where I don't actually gain anything by contributing so I just lurk hoping the schizos tire themselves out and leave. And if I were to guess I'm probably not the only one.
>>8661473I've done so much mesugaki I've now started doing more "casual gf" versions of her.
https://litter.catbox.moe/0fqswrrylmbejtgq.png
>>8661479We've also just hit a period of complete stagnation and stagnation doesn't lead to very active discussion and tends to be the point where schizos easily take over. The only thing that really keeps me personally going is that I just keep building datasets and training stuff and trying to push the limits of what can be in a dataset and still work.
>>8661482Its not the same without the twintail
>>8661543If it's one tail can you really call it twin?
>>8661465Never ever went to hdg
I love this place because of the theorycraft
>>8661602To me they're still the same place, just a pointless split.
>>8661616I'm perfectly fine not having to navigate through a shit ton of hentai gens to see useful advice.
>>8661631There are no hentai gens in that dogshit thread
Man, Chroma is slow as fuck even compared to Flux
>>8661543You're partially right, but I like alternative hairstyles. Gives more life to a character when you can change their hair/outfit/etc. but still recognize them.
https://litter.catbox.moe/b9sui6vg4puaqghb.png
>>8661639and thats after he reduced the amount of parameters. these models are designed for api, h100 minimum. flux api returns gens in less than 3 seconds. local consumer hardware is outdated. local models cannot significantly improve with the current state of hardware
>randomly check noob's civit page
>LAX liked a comment asking when the model will be updated
>no other recent comments were liked by him
bros...
>>8661639>>8661648Did you run the gguf quantized model or just the regular one? Smallest runs faster than sdxl.
>>8661656I ran one of the gguf ones, but I didn't go for the smallest since I was afraid it would completely nuke the quality
>>8661652LAX wants noob v2 to be a flux kontext-like editing model, not a direct t2i model
>>8661666they want something like omnigen 2 which is capable of both
>>8661324Reminds me of this.
so who's making a lora of the grok girl
you need a lora for that? She looks so generic she was probably genned on WAI in the first place
>>8661788>you need a lora for thatIf you want to get the clothing exactly on model, yeah.
localkeks are just so used to their model being unable to do anything they just automatically think everything needs a lora
the civit sloppers are already one it
https://civitai.com/search/models?baseModel=Illustrious&baseModel=NoobAI&sortBy=models_v9%3AcreatedAt%3Adesc&query=grok
>>8660637 (OP)I got OP spot? Wtf
>>8661799Took longer than the Steam sale girl
>>8661799I need a quality /hgg/ approved one.
>>8661826civitai loras surpassed 4chan loras back since the pony days
>>8661652Why do you still care about this NovelAI shill?
>>8661809the dick x2 spell
Why did you think you had a choice of not suffering here?
SUFFER!!!!
>>8661876only she cast it too early
>>8661880I look like this
https://litter.catbox.moe/dorzxq8an2g6ed0k.jpg
>>8661828Ogey but it can still be anon approved.
I fucking hate inpainting hair
>>8661955could have been better without the raised arm
anyone got a speech bubble png kit?
>>86606901girl,blonde hair,large breasts,naked,shibari,suspension,upside down,spitroast,cum on breasts,cum in pussy,cum dripping from mouth
>>8661478We're fucking cooked.
>>8661528I just like making loras of my gacha wives. I've peaked.
>>8662016spitroast never works for me. It's always better to just prompt oral and sex from behind.
>>8660637 (OP)Checking /hgg/ and /hdg/ , it seems like /hgg/ is maybe marginally better. Is that the general consensus?
>>8661806Nice job with the gens. How are you doing the thought bubble? Editing it in after the initial gen and then inpainting the image inside?
>>8662042Thanks for the tip. If anyone wants to try, just sub out "spitroast" for "oral, vaginal."
I am on the verge of downloading fucking reweik again just to do some more gens on some of my old styles with my current ""knowledge"" just to see how much better I can do it now
nobody is using pony in 2025, lil blud
>>8662056Thats right. Gen it with any thought bubble just to get it in the base image, then inpaint.
Rajeesh still uses the ponies, sir.
>>8661528I've moved on to video genning. Things that used to take an hour are now under a minute. It's pathetic this place doesn't even have a guide for wan.
>>8662133it still looks like atrocious garbage though? but i'm happy you're having fun with it
Anonies I just came back from a long vacation from slop, what's the current meta?
>>8662139nai 4.5 came out and will be better than local for the next 2 years confirmed by LAX
>it looks like [headcanon]
btw what lax said specifically is that the difference between neta lumina and nai 4.5 is significant but you guys can't read
>>8662144You are not getting out of this one, scumbag. Everyone has seen that you're shilling for NovelAI. Everybody knows you purposefully sabotaged NoobAI v-pred because you got paid off.
>>8660974Us localkeks will never be able to taste this with our own schizo 3d models and styles...
>>8660661Should be tentacles instead of dicks.
Miko, shibari and tentacles is ageless combination.
>>8662141Except nai doesn't have gopro lora, so in reality local is gazillion years ahead.
>>8662016"upside down" really makes few angles work on this one, but I felt removing it would be too far from your vision.
what model are you guys using? naixl2d custom? no eps pls
>>8662221Come on dude just read the archive, we do this every thread. 291h, or 102d custom. I use base v-pred too but most people don't want that recommender.
>>8662221nai 4.5
before that noob vpred 1.0 with a couple of loras
Uh oh, the creature got angry.
Uh oh the schizo is babbling about creatures in xhis head again.
>>8662133Post something. I want to learn.
So how do we all feel about "ani_(xai)?"
>>8662278If I didn't already have three twintailed sluts I was plapping I would. At this point it's pretty generic.
>>8662221There are like a million and one merges and more coming out every single day. They're all close enough that the prompt and any effort you're willing to put in to editing or inpainting or such will make a much bigger difference to the end result than the merge.
As far as I'm aware, WAI is still the best standard / the yardstick everything else is measured against. Nova models are inflexible but have solid default styles, there are a bunch of variations of Nova models if you want a specific look. MolKeun is more flexible for a wider variety of styles and 3WolfMond is also flexible for many styles including anime but does better realism with a more realistic default look.
:skull: :skull:
:skull:
:skull:
what's wrong with WAI though?
DAILY MODEL WATCH
Since people are asking and there are new merges every day might as well test them and see what they're like.
Here's 3 new ones:
https://civitai.com/models/1779741/wai-nai-hybrid-model
https://civitai.com/models/1778592/corepom-mix2
https://civitai.com/models/1778310/rougenaut
Pic related shows how they compare to WAI.
Full res test: https://files.catbox.moe/wlm5hq.png
Rougenaut is the one that stands out the most, but not in a good way. It has more of a painterly style, but seems to constrained to close-ups. Corepom has more of a retro anime look. The WAI-NAI hybrid seems to do wider shots would could be good but doesn't seem to match the quality of default NAI.
Very interesting anons around here the past days
The Studio Ghibli style LORA got enough attention to make mainstream news, but there are a lot of other LORAs for other specific anime styles too.
The most downloaded / most highly voted one for Dragonball is https://civitai.com/models/474074/dragon-ball-style . It captures the look of the original anime quite well.
If I wanted to learn about image gen, what sources would i read first?
>>8662365The OP links would probably be everything you need, but really starting from scratch and knowing absolutely nothing a good guide is https://stable-diffusion-art.com/automatic1111/
>>8662334the fact that despite getting noobai/illustrious local still settled for WAI (aka autismmix v2) as its standard model is the best proof that most of ai users are blind 90 iq subhuman pajeets
>>8662354the absolute state of local...
>>8662367I think people tend to use Illustrious or Noob merges rather than the base models. Here's how WAI compares to the base models though.
One thing history has shown time and time again is that people prefer convenience over most other things. As pic related shows while Illustrious and NoobAI may be more technically capable models, WAI looks good by default. Most people probably prefer to just click a button and get something that looks good rather than work on a prompt to fix issues and improve style and such.
>>8662386What the fuck went wrong with Noob vpred? Was eps 0.5 unironically the best release?
>>8662213OR here. No worries about it, I think it came out looking great. Thanks for working with my weird prompt combo.
>>8662391eps 0.5 shits itself without lora as well. all base noob/illustrious models pretty much require some low-weight lora.
>>8662386Of course.
It shows that who the plebs are and who the people who invest time are. Though in this case it's not even that high of a wall as downloading and testing loras which have a stabilizing effect really does not take that much time.
>>8662391Vpred models use a different noise scheduler that allows for more color depth, so Vpred models can do darker scenes better and handle more contrast with lighting and shading.
However, that makes Vpred model more finicky to use. You need to use a Zero Terminal SNR noise scheduler, a CFG Rescaler, and might need to be more specific in the prompt about lighting and such.
Vpred is technically superior which is why merges with the Vpred NoobAI model are popular, but I agree that the base model looks rough unless you really work with it.
>>8662391A same-settings comparison between eps- and v-pred does not make sense.
>>8662412This is an accurate statement, but unfortunately as far as I'm aware there's no real way to do a direct comparison between them while changing settings and such. I guess the best bet would be to use the same prompt and seed and just change other settings and then compare the results? But then that's highly seed dependent since one seed could look good on one model and not on another.
>>8662386>WAI looks good by defaultit doesn't
>>8662416WAI is consistent by default. Like McDonald's food you can decide for yourself whether you like it or not but people who go there know exactly what they're going to get so they keep going back.
Also pic related and some others I've posted are NovaAnime and not WAI, NovaAnime is my preferred "default" anime look but WAI is more popular so makes for better comparisons and such.
>>8662422yeah i guess that's a fair food analogy
>>8662422Yeah it's really good for Civitai genning when you have a limited number or rerolls/inpaints. But the style leaks into everything you do and knowledge is overall weaker.
>>8662426That's probably a good assessment. I didn't consider limited gens because I run locally, but for anyone who is under limited gen constraints having things at least turn out decently by default is probably very important.
>>8662016>>8662042This seems like something that might be model knowledge dependent. "spitroast" is an official Danbooru tag, so models trained on Danbooru data should know it. Here's a quick test where it looks like just using "spitroast" on its own worked best, but your results may vary.
are we getting raided by civitai what the fuck is this thread
>>8662477butiful extra set of vestigial eyes on the reddit frog sir
>>8662477NovelAI's v4.5 model is probably the best model currently available for uncensored anime gens. However, there's a project to train an uncensored anime model based on Flux that's currently ongoing called Chroma: https://civitai.com/models/1330309/chroma
Hopefully it'll be successful and local will have an equivalent model to NAI's v4.5.
>>8662480this is only for furry shit its gonna need an anime finetune or gorillion loras
its over
>>8662480so true sister!!
>v4.5 is anything but pure hot garbage
the only passable model they released was v3, admit it.
just another psy-op to make anons go back to hdg because the threads are "the same"
>>8662480As of right now Chroma is worse than Pony v6. And it's almost finished baking.
Look at how uppity they get over one NAI image.
>>8662386Wish there was one like WAI but with absolute style adherence
>>8662480best case scenario (model is successful, gens good shit and doesn't fall apart) it's still a primary furry model and it's gonna be the EF/Pony sepia lora hell all over again. Also add the ridiculous gen times on anything that isn't 5090 or better. Fuck that. Noob is unironically good enough to gen occasional slop and if you plan on binge genning once in a while, 4.5 is good as well. Fuck furry models.
>>8662489Are you saying shitposting doesn't actually degrade the quality of the thread? It only makes you think it does?
file
md5: d76b2d0ca030428849db33311cff7080
🔍
>>8662398Recommend loras with stabilizing effect then dipshit
I always see this argument thrown around and haven't seen a single one being shown with proper tests to assure it's efficacy
Models trained with more parameters have more knowledge and capabilities than models trained with less parameters. Right now Flux is the publicly available with the most parameters, but the volume makes training prohibitively expensive and resource intensive. Unfortunately without monetization it's hard for anyone to justify the cost of training.
>>8662490It's only at v0.41 current though. I presume the intention is to make 1.0 the initial release build so it seems like there's more than halfway to go for that yet.
>>8662493>Also add the ridiculous gen times on anything that isn't 5090 or better.I've only got a 3090 but I've run Flux and though it takes longer than SDXL it's not that bad. Maybe a minute or couple per gen? It's been a while since I've tried since that was with the base Flux model which is censored but it was impressive to see the prompt adherence and how it could generate legible text and such.
>>8662499it's 41 epochs out of 50
>>8662499>Maybe a minute or couple per gen?yeah that's ridiculous on a 3090 wtf?
Wish I could run flux locally
it's over 3050 8gb vram bros
>>8662499>>8662500It's actually at 44/50 but 41 wasn't really noticeably better than 31 so why care
>>8662492What kind of styles are you trying? Pic related is a test showing how some models respond to different Danbooru style tags, some professional anime artists, and the top artist tag on Danbooru.
It seems like all of them are at least fairly responsive, even without LORAs.
Full res test: https://files.catbox.moe/y2fcp8.png
>>8662501Again, it's been a while so I could be misremembering. But it's definitely shorter than the video gens I've done at least because I didn't like walk away from the computer and do other things while it ran. So it might take multiple minutes but not like hours or such.
>>8662497Not that anon, and I don't think that "stabilizer" LORAs consistently improve gens, but here's a test of some LORAs available using WAI as the model.
https://civitai.com/models/971952/stabilizer-ilnai is the first result that comes up when searching for "stabilizer." It's the most popular/most downloaded/highest rated. It just says to use it at one weight and that it should make gens better. This one just seems to add realism to a gen? Especially in the first example. Up to you if that's what you want I guess.
https://civitai.com/models/99619/control-lora-collection or NAI Vpred Fix is one of the first LORAs to come up if you just search for NoobAI LORAs. The description says to start using it at half strength so I tested with both half and full here. It definitely does SOMETHING, though whether it makes gens better or worse is your call.
https://civitai.com/models/1001945/detail-slider-lora-or-illustrious-xl is a detail slider that's supposed to add more details to gens. The examples showed using it at 3 weight so I tested it with both 1 and 3 weight here. This does seem to succeed in adding more details, though whether that's the look you want for a gen or not is your call.
Full res test: https://files.catbox.moe/pkptdl.png
>>8662502That's enough to run the gguf versions
>>8662497Any decent style loras of your preference and especially mixed bro. Not going to put together a collage, too lazy to set that up in Comfy.
>>8662517Generally those loras are bad. What we usually mean by stabilizing loras are just good style loras which also happen to make the model less prone to generating garbage. You need to test it paired with the base model since shitmixes already are stable with exceptions.
>>8662480>NovelAI surpasses localslop with a new model>Local turns to furshit to copethis cycle has happened 3 times now
>>8662526>What we usually mean by stabilizing loras are just good style loras which also happen to make the model less prone to generating garbage.Do you have any examples? It'd be neat to try them out!
>>8662544can your next bit be the one where you sub to nai and refuse to use artist tags again? it'd be especially funny with v4.5 since it looks so dogshit without them
>>8662544Yeah I think this guy's loras aren't too bad.
https://civitai.com/models/301980/merrytail-style-noob-vpredillustriouspony
I might be wrong though since I use a mix of like 8 different loras at lowish weight that combine to fit my style preferences so I don't know if individually they're good or not.
>>8662507A lot of styles
Mostly replicating actual artist styles (generally towards the more realistic/painterly side) or trying to make something that looks like an actual screencap from a show
The slop that goes into these models usually ruins both lol
>>8662526Yeah that's super vague and as usual the fags that say it don't give any sensible proofs
Either get your lazy ass to run a proper test or recommend loras that are failproof so that other people can test it
Otherwise you retards will just go "uhhh that one doesn't count"
>>8662549Thanks for the link! If prompting can accomplish something I tend to prefer doing that then using a LORA, but for styles LORAs can often have a far greater impact.
Pic related is a quick test using WAI of the LORA you linked to plus another from the same creator that was called "best in one" so I thought it was worth checking out: https://civitai.com/models/723360/ai-styles-dump-illustriousrouweinoob
no matter the guidance i always burn the image in nai and the characters have weird anatomy and faces, i need the face detailer omg. probably because i'm not used to it yet
>>8662551For making an image looks like a screenshot, the tags I'm aware of are anime screenshot (official Danbooru tag), anime screencap (unofficial but seems to help with some models), and anime coloring (Danbooru aliases "cel shading" to this).
Is there anything else that could make an image look more screenshot-y than that, other than subtitles?
>>8662556>face detailerUse focused inpainting. It's the third tool on their inpaint UI. Or use the NAI extension.
>>8662565anime_coloring, anime_screenshot, official_art, on_model, sometimes film_grain and jaggy_lines help
Also noob models are trained on e621 too so using their tags help as well
>>8662568Thanks for the heads up! Here's the same shot as
>>8662565 with more of those added, so the style tags are "anime screenshot, anime screencap, anime coloring, official art, on model, film grain,"
>>8662565>anime screencapI believe danbooru renamed anime screenshot to anime screencap and then back at some point, so it depends on when the model was trained.
>>8662568And of course, the name of the show you're trying to mimic
It helps to add tags that often show up in the style
For instance dragon ball is pretty old so retro_artstyle and 1990s_\(style\) helps
AOT often has warm_lighting and highlight_\(coloring\)
>>8662572That did make it better
>>8662573Oh, I did not know that. Thanks, that makes a lot of sense.
>>8662574Ahah, that's good thinking! The only issue there is you'd need to know the tags and change them for each series you're trying to emulate. Still for
>>8662551 it seems like you could get a long way towards replicating a show's look even without using LORAs if you used all the right tags.
Here's a quick test of those combining them with the previous tags, though like you said it'd really be a matter of picking and choosing the right ones to make a specific show's style.
>>8662584Vpred sucks but it's absolutely superior in this aspect when compared to mixes and all
I'd be fine with a model losing knowledge for consistency outside of artist and styling tags
>>8662270Oh, forgot to setup the stealth png metadata I guess. Here you go:
<https://litter.catbox.moe/ub80la2ns15l5o3b.png>
>>8662587There are some mixes geared in that direction. For instance NovaAnime is very consistent but sacrifices style flexibility for that consistency: https://civitai.com/models/376130/nova-anime-xl
There are a bunch of different Nova models all with different built in styles because of that design philosophy, each different style basically needs a different version of the model because they'll all built in so strongly. But if you like the look of any of them you can reliably get that look without any extra tags or anything.
>>8662521catbox? nice style
>>8662606>sacrifices style flexibility for that consistencyThen it's the opposite of what I want anon
I want a consistent model with great style flexibility and adherence
I don't mind losing knowledge in other areas so long for the consistency but all of them seem to fuck up styling tags
>>8662626Oh sorry, I misunderstand. Yeah, unfortunately I don't think there's any models that focus on styles and such so much they sacrifice knowledge in other areas. LORAs are probably the best way to add more style flexibility and adherence to models.
Tags that might be underused include things like "symmetry" and "rotational symmetry." The former tries to mirror and image and the latter does a rotational mirror like a yin-yang.
Pic related is a test showing off the effects.
It seems like these kind of tags would be unlikely for people to stumble across, you just already have to know they exist and what they do. Does anyone else know any tags like that that people might not know and would be hard to figure out?
>>8662653I do have a list of tags (both niche and common) that I like.
>>8662634Since you're doing these cool comparisons, would you mind testing the prompt adherence/style flexibility of these? I've been wanting to try them for a while
https://civitai.com/models/934492
https://civitai.com/models/941345
civitai.com/models/1263515
Would be cool to test out a single anime screenshot and then a few artists that deviate from the standard anime style (try: yuming li or mgong520)
Also, for the "no style" gen, you can just leave the last tag with a comma and it'll automatically gen one without the tags
Like this: tag1, tag2, tag3,
You can also combine sections of the prompt by leaving no comma between the quotes
Like this: tag1,"tag2, tag3", tag4,
>>8662653The tag contrapposto adds a dynamic element to most poses, although it does make it look effeminate
>>8662667Sure! Downloading the models now and will test them out when it's done.
For the no comma trick that's good to know, but I think putting a header makes the chart more readable and doesn't affect the gen significantly.
For the second one what do you mean? I know that like you can use multiple tags if you don't use the comma to create a new column, like in
>>8662507 where I used "retro artstyle 1980's \(style\)" like that. Is that what you mean, or something different?
>>8662671Ahah! That's another good example, good thinking anon. I was using "cocked hips" for a while for that kind of pose because I didn't know about the "contrapposto" tag.
>>8662667Here's the test of those models compared to WAI.
Included the anime screenshot look using all the tags from earlier in the thread, both the artists you mentioned, and a few other styles too.
Full res test at: https://files.catbox.moe/qxamot.png
It seems like all of them are perfectly viable mixes. MadlyMix stands out with the most distinct look of the bunch, while Hoseki seems the closest to WAI. RichyRichMix looks the most 2.5D / realistic.
I bet you could get good results with any of them though.
>>8662689Well that's all you can expect from mixes, style bias. The main difference is gonna be in how much knowledge they lose of low image count artists and characters, and how much they are overfit on common poses.
>>8662671Also thanks to contrapposto anon here's a random selection of poses from Danbooru's posture page: https://danbooru.donmai.us/wiki_pages/tag_group%3Aposture
Picked out a handful with names that didn't seem like things people would stumble across or just randomly use if they didn't already know the tags.
>>8662695Yeah, that's about my understanding as well. With models having gotten good enough and mixes being so similar, prompting and effort in editing and such makes a lot more difference in the quality of the gen than the specific mix used.
One unfortunate thing about training on Danbooru and e621 datasets is it seems like a lot of knowledge of professional artists from the original SDXL model was drowned out or lost. However, it does seem like some professional anime / manga artists still work.
Is there a list anywhere of what professional artists still work with IllustriousXL or NoobAI based models?
>>8662677>>8662689MadlyMix notably gets most styles better than the others, except the screenshot
Thanks for testing anon!
>>8662706Np, btw there's one like this for hand gestures too
I often use fighting_stance, claw_pose and clenched_hand for the traditional karate vibe
As for other tags that people wouldn't know, I think genderswap_\(ftm\) is pretty interesting
If you're doing an androgynous male character and there's a few tags that overlap too much with 1girl, your character will tend to become a woman
This tag fixes this, specially on 1.3~1.5 weight
>>8662719That is indeed a good way to create a fighting stance. Good thinking anon!
When you use tags, do you use them with or without the underscores?
>>8662756She’s so cute! I love her!
>>8662763Thanks, glad you like her! It's using the Dragonball LORA mentioned in
>>8662361
One useful extension to have is dynamic prompts: https://github.com/adieyal/sd-dynamic-prompts
This lets you replace parts of a prompt string with other strings, like {smile|frown} making the gen choose between smile and frown and having a lot of other syntax for other effects like that.
It also allows prompts to pull from wildcard files, so you can make a list of things like characters, outfits, locations, etc and the prompt can pull from the file.
I'm making a list of the most popular Pokemon girls. My current list is: https://files.catbox.moe/2fykhr.txt
Did I miss any major popular ones? There are over 100+ characters listed on https://danbooru.donmai.us/wiki_pages/list_of_pokemon_characters and most of them have too few uploads to be known by models.
Also, is Serene recognizable without her hat? Most Pokemon girl generate with their hats even if it's not prompted, but she doesn't. I want the wildcard to be flexible but I also want the characters to be recognizable. I'm leaning towards the hat being a defining part of her character design so she's not recognizable without it.
your honest reaction to everything posted this day, /hgg/?
>>8662354I like this idea. I too used to test new slops out of curiosity. It's how I stumbled unto NNAillous back in the day. Was really good for an eps but then the baker started incorporating ntrmix into the bake and that was that.
>>8662544748cmsdxl or try_style at 0.2 weight are generally good stabilizers. Although recently I merged super sekret 291 with 3 greeks at 60/40 (variation of what some anon was shilling) and I'm now lora less.
>>8662667I've tested those before. Nothing that really makes it stand out from the rest of the similar slopmixes.
>>8662773We're SO back, /hdg/sissies.
>>8662768Right might have gotten the hat but it got everything else wrong.
>>8662773Maybe we should contribute more high-effort posts...
>>8662780>I'm now lora lessWhat do loras have to do with it? I'm using loras because the model doesn't know the artist.
>>8662782I was doing that but you n-words stopped so I also did
>>8662783>What do loras have to do with itI was using low weight loras to stabilize gens. I don't use loras otherwise as I prefer mixing styles through innate model knowledge. After pony, I don't wanna hear about loras.
>>8662780>try_styleWhere's that?
Here's Aisha ClanClan of the Ctarl Ctarl empire showing off all the different swimsuit type tags on Danbooru's swimsuit page: https://danbooru.donmai.us/wiki_pages/swimsuit
This doesn't include the color or extra tags, but those could be mixed and matches with these for probably literally hundreds of combinations.
Individuals at: https://imgbox.com/g/TDNByHEB1O
I'm using the stealth metadata extension https://github.com/neggles/sd-webui-stealth-pnginfo so the prompts for these should still be preserved in the alpha channel, you'll just need that extension or another alpha channel metadata viewer to see it.
>>8662780Thanks for the info, appreciate it. I haven't used NNAillous before, is there a version that's still worth checking out?
minthy did something interesting
https://huggingface.co/Minthy/RouWei-Gemma
>>8662788Here's the hash: b39e51de3a02
>>8662789>I haven't used NNAillous before, is there a version that's still worth checking out?Nah. Outdated knowledge at this point as the good ones were still prior to noob 1.0 I believe. It was my go to when personalmixnoob was making its rounds as the go to slop.
>>8662791I was looking at this earlier. Reading the model card it's only available as a ComfyUI node, correct?
I have ComfyUI installed but only really use it for video gens. For image gens I use Forge or Reforge. I've also heard that InvokeAI apparently has a lot of capabilities out of the box, like creating a larger canvas and defining bounding boxes for gens so you can gen comic pages.
What's the /hgg/ recommended imagegen UI?
>>8662791hmm... interesting concept, wishing him good luck, at least one person is doing something
One last test for the night before bed, this time demonstrating the effects of the "flowing hair" tag at different weights.
Including this tag might be a good way to make scenes feel more dynamic or give more of a sense of motion like hair blowing in the wind.
>>8662620https://files.catbox.moe/048mvh.png
>Gaia Token Petal Alchemist Dress woven from vibrant flower petals, hair adorned with blooming vines, carries baskets of crushed blossoms, skin dusted with pollen, smells of spring
>Visage of Vengeance Petal Alchemist Dress woven from vibrant flower petals, hair adorned with blooming vines, carries baskets of crushed blossoms, skin dusted with pollen, smells of spring
I wonder what about these random AI generated prompts made these two generations spawn a penis
>>8662884https://files.catbox.moe/8eny73.jpg
>style lora I want also has a detrimental effect on certain body proportions/shape that I don't want
aaaaaaaaaaaaaa
>>8662893use lorablockweights to fix it
>>8662893I worded that a bit incompletely. The situation is that the style lora includes some aspects that I want in terms of proportions and shapes, but some aspects not. This makes it so that scheduling the lora doesn't work as it also makes the desired aspects of the lora weak.
Shit the captcha failed and delayed this post.
>>8662894Is that a training thing?
>>8662895When training it, or when using it. Wouldn't recommend it, it's a very deep rabbit hole.
Make sure you tag your dataset with bodytype tags, and add some loli in there. I know you're one of those fags.
>>8662895you can also just train on some blocks during training but no
https://github.com/hako-mikan/sd-webui-lora-block-weight
i don't know if this has been updated to the new forges (but it has flux support so it should be) it was pretty popular in the 1.5 days
you can just try to maximize the style while minimizing the effect on anatomy and posing
>>8662899If anything it's the opposite direction I want to go kek.
>>8662900Oh hmm. I guess I'll read a bit on that before I judge whether it's too deep a hole.
>>8662878>>8662885I think it's more about the model you are using being fried and overfit on porn.
>>8662915>Model: pornmasterPro_noobV18VAEgood guess
>>8662915>>8662917kek I'll keep that in mind when I'm on another model. I very rarely get any porn randomly like that though and most models are probably horny.
>>8662917Based master of porno
>>8662917I remember shilling that model before, as a utility to help with img2img between anime and photorealism. Cause it's meant to be capable of both, but instead ends up in the weird limbo between them.
https://litter.catbox.moe/dutc208051kv1xw1.jpg
asanagi becomes 2.5D instead of flat, photo gets huge eyes and airbrushed skin. Both make style changes easier.
>>8662936Periscope dick, what the hell
Which version would you recommend, if any? Seems like it could be fun to prompt artists or try loras on it
>>8662938I settled on V15 previously, but didn't try any that came afterwards.
>>8662939Alright, guess I'll try the latest and that one then, thanks
>>8662936OWWWWWWWWWW MY DICK!
>>8662942is this a femanon falseflag
there should be no pain when bending upwards, even at full hardness
>>8662949I don't even want to attempt to prove you wrong.
>>8662949this
also put bengay on your balls it feels amazing
>>8662900>>8662902I always wondered why people just didn't do block weights to remove inherent style instead of the copier method. Since reforge has it integrated you could just train a character lora, check each block manually in reforge for the one that has the style, then rebake with that one blocked and you're done.
>>8662959Because the blocks are not cleanly separated between topics, you have to mess around with it and test which block affects what and to what degree.
And since you're rebaking anyway, it's not any less effort than copier.
>>8662962>And since you're rebaking anyway, it's not any less effort than copier.Come on man this isn't true at all and you know it. For copier you have to make an entirely separate dataset for a style lora and train that first (and properly) before you train the character lora. Meanwhile with this all you really need is to figure out which blocks do what and it's over. Even if you didn't remove 100% of the style with one block it's still pretty good.
>>8662957You're just scared of the truth.
>>8662969I didn't mean leaving some partial style behind, but losing some character knowledge because you took out a block that had 4% of her clothing detail along with all that style bias.
>>8662949thanks, I came mustard gas
>>8662979That sounds like a superhero ability. You could fight crime along with a cute sidekick specialized in handjobs.
>misa lora
i didn't realize elon posted on /hgg/
Well trying to go back to reweik was clearly a mistake
DAILY MODEL WATCH
Some interesting new mix models this time around.
beret mix Manga Color - https://civitai.com/models/1783367/beret-mix-manga-color - is intended to create a colored manga look and is made by a Japanese model creator.
Karin Mix - https://civitai.com/models/1781584/karin-mix - Seems to be going for a more typical anime look and is close to WAI. Maybe more cutesy based on some of the example images on the CivitAI page.
MAM Max Manwha - https://civitai.com/models/1777051/mam-max-manhwa - is notable for going for a Manwha look instead of a standard anime one. It's probably the most distinctive of the bunch, and seems to handle dynamic action well as shown in the bottom right corner here.
Pic related is a test of all these compared to WAI as the control group.
Full res test: https://files.catbox.moe/olz396.png
Feel free to let me know if anyone has any models they'd want to see tested like this!
Does anyone know if there's a way to search Danbooru for characters matching a specific tag? I'm trying to get a list of kemonomimi characters but there doesn't seem to be any way to narrow down the list of character tags at https://danbooru.donmai.us/tags?commit=Search&search%5Bcategory%5D=4&search%5Bhide_empty%5D=yes&search%5Border%5D=count
>>8663070Related Tags tab on the search page
https://danbooru.donmai.us/related_tag?commit=Search&search%5Bcategory%5D=Character&search%5Border%5D=Overlap&search%5Bquery%5D=cat_ears+
This gives you a list of characters ordered by how often their images include cat_ears.
>>8663076You are awesome anon! Thanks a lot for this, really appreciate it!
Does Noob have an issue generating nipple rings under clothes? Trying to do something like this.
https://danbooru.donmai.us/posts/9390597
But when I test something like "breast curtains, nipple chain, covered piercing, nipple rings, see-through clothes", it makes the nipple rings go above the clothes. This is good if I wanted "piercing_through_clothes" but I don't. I also tried variations like "covered piercing nipple rings", "covered piercing, covered nipple rings", "covered nipple rings", and some prompt editing attempts, but none of them worked.
>>8663107covered piercing is both too ambiguous and lacking enough pictures for noob to learn properly, you'd probably need to train a lora for it.
>>8663107try prompt scheduling
>>8663107The anon in
>>8663115 might be a genius or I might be an idiot because I thought he was talking about https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing so that's what I used and got it to work.
This prompt editing lets you have the prompt switch from one string to another at a certain step in the generation. I used "nipple rings" and "covered nipples" and tried switching at each step until one worked. For pic related it looks like step 20 was the sweet spot, and the only step where it actually worked.
You can see below 20 the nipple rings didn't gen and above 20 the nipple rings are outside the clothes. Step 20 actually got the ring under the clothes. The gen is using 30 steps total so the switch is 2/3rds the way through.
The prompt part for the switch was [nipple rings,:covered nipples,:20] . You'd probably need to tweak the steps for each gen until you found one that worked though. You could increase the total number of steps in the gen to have more precision when switching too.
Hopefully this is the kind of thing Flux based models could do a lot better if they have a better understanding of in front / behind concepts.
>>8663130I did mean prompt scheduling, it's the same thing. But it looks pretty rough there. Inpainting is another solution as well. Anyway just wanted to confirm this wasn't some other issue on my end or a model issue.
>>8663138Ah, that makes sense. And yeah editing / inpainting is usually the best way to get the results you want for anything, it's just a matter of investing the effort to do it.
Glad you brought up prompt scheduling though because even though I know these techniques exist I always forget to ever use them for anything.
Does anyone have a common use case where they find prompt scheduling or other kind of techniques to be really useful?
>>8663139Yeah. I use prompt editing with certain artists that result in a body type I like, but since the style isn't to my taste, I switch to a different artist mix at around 0.4. I use it often for making things closer to a specific shade of color that I want (in conjunction with negatives). I use it sometimes for making certain types of clothing designs that aren't normally generated. For instance, a one-piece swimsuit where the top is a specific different color from the bottom, or a different shape that resembles other clothing.
>>8663150Oh wow, that is way more applicability than I was thinking. Thanks for the explanation! I guess I've just never thought to try it for things like that. I'll have to try experimenting with it more often!
Wow, they attached a personal friendly LLM to /hgg/, and it even posts images!! Very cool. I think whoever created it needs to work on it's imagegen skills though... Not only the images are mid, most of them are off-topic for /h/.
did anyone ever make an updated, decent lora training guide for illustrious/noob-based models? i still havent been able to make any good loras (not that i have tried all that much)
>>8663179It's not like you need one, I've baked decent noob loras with settings from 2 years ago. The requirements haven't changed, we're still on SDXL and even SD1.5 wasn't that different. Just gotta tick v-param and zero-term-snr if you're training on noob, and that's it.
What the new stuff gives you is mostly speedup. Gradient accumulation, fp8, faster optimizers. I read some people now bake in 300 steps what would have taken 2-3K back then.
are cuda malloc and cuda stream worth using? can't even tell if they're helping, left is default, right is with both enabled. 3080 12gb. simple 4 gens 832x1216
>>8663179I have not seen an updated guide, the one I rely on is still: https://civitai.com/articles/138/making-a-lora-is-like-baking-a-cake
However, an anon in another thread did upload his .json configuration for training on IllustriousXL/NoobAI: https://files.catbox.moe/h572ls.json
If you're using koyha_ss then you should just be able to open that configuration file, otherwise you could open it in any text viewer and see the settings and set them in whatever trainer you're using.
It seems like /hgg/ does a lot more LORA training than other threads. Has anyone figured out optimized settings for training on IllustriousXL or NoobAI?
>>8663184>>8663186thanks, i will try again later
>>8663186Check previous /hgg/ threads, feels like in every other one someone shares a .toml with their settings.
>>8663188I'll do that, thanks!
This webpage appears to be a discussion thread on a community forum, specifically focused on AI-generated content, known as "local" or "stable diffusion." The thread is centered around the topic of disappointment with the current state of local, with users discussing its limitations and comparing it to previous versions. The conversation involves users sharing their experiences with local, discussing the quality of generated images, and debating the effectiveness of various techniques and tools. The thread also touches on the topic of AI advancements and the challenges of maintaining high-quality local content.
>>8663202I mean, it's not wrong tbqh
At least as far as I can tell, here are the most recent LORA training settings posted in the thread as .toml files:
https://files.catbox.moe/1zxhfw.toml
https://files.catbox.moe/g26bov.toml
https://files.catbox.moe/9u9jei.toml
https://files.catbox.moe/3kx4ms.toml
https://files.catbox.moe/72zp6j.toml
Looking at all of them, it seems like the vary a lot. Is there no general consensus on what works the best?
Just in regards to the optimizer, one uses AdamW, one uses AdamW8bit, two use Came (which I had to look up because it's not in my Kohya install by default, it's https://github.com/yangluo7/CAME/ ), and one uses something called "torchastic.Compass" which I couldn't even find trying to look it up.
It was my impression that AdamW was basically the default option, AdamW8bit was smaller if you needed faster training or less memory, and that algorithms which can change the learning rate dynamically like Prodigy were supposed to deliver better results.
Is that not the case, or are there other considerations? I see from https://github.com/yangluo7/CAME/ that CAME seems to be optimized for speed and lower memory usage, so is that just what people are valuing more?
>>8663229>torchastic.CompassThat's a finetune config, not a lora one.
>>8663231Ah, that explains why the formatting is different too! Thanks for that, saw that it was different but couldn't tell why at a glance.
>>8663229Dataset matters way more than settings, as long as they're not outright "wrong" like using noise offset with v-pred, or a crazy high learning rate, etc.
>Prodigy were supposed to deliver better resultsNot necessarily better, just save you fiddling with the learning rate. After a few steps it's supposed to figure out a decent LR on its own. But even with Adam you don't usually have to change LR from the default, so I think it's kinda pointless and a potential point of failure.
CAME with rex warm-annealing restarts is what anons mentioned with the 300 steps thing. Together with gradient accumulation which simulates a higher batch size, it's supposed to train very fast.
>>8663246Thanks for the info.
>Dataset matters way more than settings,Ah, damn. Prepping a dataset is the most pain in the ass part of training a LORA. It seems like you need to do a lot of manual tagging and image editing if you want a really good dataset.
>CAME with rex warm-annealing restarts is what anons mentioned with the 300 steps thing. Together with gradient accumulation which simulates a higher batch size, it's supposed to train very fast.That is good to know! I guess I was under the impression that something like Prodigy would have a lot more value than it apparently does. I've also gotten results with AdamW that were workable at least
Maybe I'll have to try using CAME instead to see how much of a difference it makes. Never would have found out about it if not for /hgg/ , so thanks to the anons who still actually hang out here and make it worthwhile!
Is Noob capable of understanding the layered_bikini or layered_swimsuit tags?
>>8663277Just tried it out a bit more. Results seemed inconsistent but it can be done if you prompt for each different piece of clothing and get the right seed. Depends on luck a bit.
>>8663277Here's a quick test showing the "layered swimsuit" and "layered bikini" tags on IllustriousXL and NoobAI (plus merges based on them too).
Looks like it works to me! Though the visibility of the layers can vary, you could try playing around with the weights depending on what look you're going for.
Is noob capable of understanding the imminent death, imminent suicide and imminent bestiality tags?
>>8662830You can also prompt wind and floating hair
>>8663069Cool tests as usual anon
>>8663283>>8663277Okay, I got thrown off by the first column working well, the anon in
>>8663281 is totally right that it's much more inconsistent than I made it seem. It can certainly still work though, here's another quick grid.
ever since switching to forge ui and illustrious, when i inpaint the border seams are so obvious even at low denoise values. that border is .25 with the rest being the default settings. I've tried essentially every combination of padding and mask blur and it still glaringly obvious. When i increase the padding the inpainted part is noticeably more blurry and I can still see the seams. Anyone have tips to make it more seamless?
>>8663286Good point anon! Here's a quick test to show the different effect of all the tags side by side.
>>8663297Seconding this question because I've never found a good solution for inpainting halos either. Sometimes using another Sampler like DDIM can help a bit. And I crank padding and blur up to max, and use the "Soft Inpainting" option in ReForge, but still sometimes the edits are obvious.
Soooooooo no one is telling them anything at all out of mere ""amability""? or are you just waiting to see if they leave by themselves?
O-ohhh I mean
:skull: :skull:
>>8663297try using soft inpainting and increase the mask blur to 32
>>8663297If you're using vpred don't even bother
Goddamn I hate that vpred is the best model someone please make a less bad one
>gens will have one hand with 5 fingers, looks great, but then another hand with only 4 counting the thumb
This is consistently happening, what is causing this?
>>8663346Unfortunately hands are a struggle for the AI because there are so many different possible hand positions and configurations and they're largely untagged.
Are you using a model with e621 data? If so there they have tags for both "4 fingers" and "5 fingers" like that, so sometimes using those tags can help.
>>8663339I don't have this issue on base v-pred. Using comfy though if that matters.
>>8663069>Japanese model creator>most slopped of the bunchIt's so over for anime. Manwha looks the most aesthetically pleasing of the bunch.
>>8663346Shouldn't be a consistent issue with noob or even the most slopped shitmixes of it. Hands aren't 100 or anything but still a problem of the past. Could be your artist mix (sometimes styles with little line art can make the model sperg out) or you're prompting some complicated composition. Try prompting for the hands to do something like a grab or reach or etc.
>>8663346Can also be caused by a bad lora, especially stuff trained on AI art. I have a pair that mess up hands much more often than the model does on its own.
>>8663365Manwha definitely stands out among dozens of models I've looked at for being distinct at least. There are so many anime focused models that are so close together it's hard to tell if there are any benefits of using one over the other. At least when a model goes for something different like manwha it can have a niche.
>>8663069Now run the same thing without quality tags and negs. They carry enough style bias to iron out most of the differences between models.
>>8663382That's actually a very good point!
However, if people actually use the models aren't they going to use them with the quality tags? Although you're right that excluding them would show the model differences more, I feel like it's a more realistic test to include them since that's how people would likely actually use the models.
What does /hgg/ think?
>>8663386Yeah guess everyone does things differently. I keep in masterpiece out of habit but not the others, and use no negs by default unless there's something specific that needs removing. Most artist prompts provide a similar effect to quality tags anyway.
Posted in the wrong thread
How do I prompt that one meme where a bunch of dudes stand behind a girl sitting (or laying) on a couch? The dudes either stand in front of the couch, or are sitting on the couch next to the girl, or there's no girl at all.
The masks are properly set, and were working a few hours ago.
>>8663386Quality tags suck, but you do have a point
However I personally will never ever use them
>>8662444Still, it has a high degree of failure, so it's up to RNG
>>8663393I believe the tag for what you mean is "piper perri surrounded \(meme\)" on Danbooru and just "piper perri surrounded" on e621. Depending on what model you're using and how it was trained you could use one or the other or both. You could also include supporting tags like "nervous" and "sitting on couch."
However, without something like Regional Prompter it's probably going to be hard to prompt for separate guys in the background apart from the main character.
Pic related is a quick test with a model called 3WolfMondAI.safetensors [36e750cdd9] since it has good prompt comprehension. You can see the last version is the one that got the pose the best.
>>8663394>>8663390At least for NoobAI based models, I think the quality tags do make a genuine consistent difference.
There's an explanation of the NoobAI quality tags at https://huggingface.co/Laxhar/noobai-XL-Vpred-0.5?not-for-all-audiences=true . One of the algorithms the aesthetic tags used was Waifu-Scorer, and that algorithm uses this example image to demonstrate the scoring system: https://files.catbox.moe/9jlipg.png
Pic related is a test using noobaiXLNAIXL_vPred10Version.safetensors [ea349eeae8] . I took off all the positive and negative tags, then added the top tags from both the Quality and Aesthetics sections to demonstrate their individual effects, then added all the current quality tags I use combined.
I think it shows the quality tags really do make a difference.
Now they definitely impact the style of the image and if that's not the style influence you want then it can make sense to exclude them, but I think for general purpose gens they really improve consistent quality.
Full res test: https://files.catbox.moe/vr4zuj.png
>>8663401>something like Regional PrompterI've been using masks, separating the image in two parts: bottom square should be the girl on the couch (and the couch) and the top square the dudes standing behind.
This arrangement fails a lot, but it gives me some control over what dudes I get
>>8663409Do you think pinup style has something to do with this result. I've always done no negs because it gets the most out of whatever I prompted.
>>8663487I stole the workflow and I don't know how to use half the nodes. https://files.catbox.moe/prxyxr.png
>>8663442Ah, that was just trying to get a more interesting pose. Here's the test again without that tag.
Full res test: https://files.catbox.moe/j3lq5z.png
What's the settings for cn tile
file
md5: 44f9554f4684e3a2f05ef7c5d27f3247
🔍
>spend my free time over 3 days or so working on an admittedly simple titty fuck image
>All because I couldn't generate a squeeze/misaligned set of titties how I wanted them
>and because it was a request by someone
Well that was silly, now it's popping up every time no issue.
How was your week?
>>8663496>misaligned set of tittiesbox?
>>8663352>>8663365>>8663376Thanks to all of you, I think it was a combo of things. Specific tags that imply use of hands help a lot, and also the reverse where it probably becomes confused about holding certain objects.
>>8663496>How was your week?Full of /e/ gens, just like the past one, and the one before, and the one before that one
>>8663492This does prove I should use masterpiece I suppose
Doesn't seem to affect the style very much and makes it much better
Does it work well for 2.5d, realistic or painterly styles though?
>>8663497Sorry I went to get food
https://files.catbox.moe/h0lili.png
>>8663513Oh it was as simple as unaligned breasts. Thank you.
>>8663513>year 2023That actually helps?
>>8663492I only use masterpiece, best quality and newest
I've found that they influence style the least
Using too many quality tags does screw up a lot of styles
>>8663516Depends on the model, actually.
>>8663297This looks a lot like the guy from Rapunzel. Intentional?
>>8663409Now add an artist or mix to the X axis. Doesn't it also make gens "higher quality"? And I think nobody uses illu/noob with the default style, it's always one or more artists or a lora.
imo quality tags are just a style bias, it's impossible to train the concept of a general "quality" so they instead push gens away from greyscale, flat color, chibi or extreme proportions, depth of field/blur, etc. and towards the "ai-generated" look. When those things might be exactly what you're trying to do, and even if they aren't it changes your artist's style. Usually in a good way, but not if they're already shiny and 2.5D.
>>8663660The girl looks exactly like https://rule34.xxx/index.php?page=post&s=view&id=7102040
Intentional?
>>8663409>NoPos NoNegSoulful
Shoo. Stinky cat. Containment thread for you. Shoo.
I like how he said baking and then did nothing for an hour. That's some creative trolling.
>>8663703I'm still waiting for the new thread before I post
>>8663696Why the third world lingo?
I'd make the next thread but there isn't a single post in the last thread worth using for the OP
gommenasorry
>>8663731In this thread, you mean? Well it's only a hundred pics, minus the XY charts.
>>8663735was talking more for hdg. though this thread needs to be remade now, too
![0628_[vocaloid]_IA_004](/images/h/1752512290986478.jpg)
