Explosion edition
Previous Thread
>>8630793>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
If you read this, have a good day 'nonie. Happy genning, happy training.
New lumina model is out
https://huggingface.co/neta-art/Neta-Lumina
https://nieta-art.feishu.cn/wiki/RVBgwvzBqiCvQ7kOMm1cM6NdnNc
>>8639670I will have a bad and angry day :( grr
>>8639673haven't downloaded it yet so i don't know
file
md5: 9c80cc1100a6c091e72e69a8a56489c8
๐
looks like it can do some suggestive stuff
>>8639670Stop, I don't know how to react to people being nice to me.
>>8639673The recommended prompt format is so cringe I don't even want to try it.
https://nieta-art.feishu.cn/wiki/RVBgwvzBqiCvQ7kOMm1cM6NdnNc#EjzAdrebEotpkKx6qlxch1XRn0g
>>8639678If you want it less blurry, the model can supposedly handle the full CFG range the way base SDXL models do. Won't burn until around 20.
>only able to scrounge together 19 decent images for a character lora
surely this wont failbake...
>>8639719characters are not that bad
i was getting half decent stuff on 10 image tests
Here's my autistic never-ever request again:
Requesting 18 sexy/cool pregnant pokegirls (or PokeMikus) crackling in their types' energy ala
>>>/vp/57920449
>>>/vp/57920452
>>>/vp/57920459
I want the bellies to look super lewd and erotic and the girls cool and radiating power
A /d/ anon >>>/d/11315717 delivered on this /r/ (https://files.catbox.moe/zet8d6.zip) but some don't quite thematically match what I'm looking for. He said he was happy if I let someone else have a go as he didn't really try to match my /vp/ posts all that much. Looking for someone who likes my idea enough to tolerate my detailed picky criteria and who likes the back-and-forth.
If you don't mind me being a stickler I have my critiques of his gens, and what I'm looking for, here: https://rentry.co/25bqq578
>>8639727would you mind sharing your settings?
>>8639730Pregnancy is such a weird fetish. Is it something you develop after turning 30?
>>863973733 so I guess. I've had it since 20 though.
Maybe this'll convert you: https://exhentai.org/g/2421726/896ce766b7/
pregnancy is worse than toddlercon
they want to fuck the baby before it's even out of the womb
file
md5: 0edd947b9d391f10d03065888e67c3ef
๐
What version of Wai do you guys use? I know it has its problems but everyone talks about it/uses it so much that I wanted to give it a go. Except there are 14 fucking versions.
>>8639749Non-pregfags assume this a lot avout the pregnancy fetish but it isn't true.
A. sticking your dick inside a cauldron of motherhood is just diamonds inducing
B. pregnant women get intermitent bouts of extreme horiness and that's hot
C. pregnant sex being somehow harmful for the baby is an ancient myth. In fact the inverse is true: pregnant sex is actually HEALTHIER than abstaining. You do a potential mother a service by dicking her down. Gently of course.
D. feeling the weight of the results of your own virility on your dick is great, pregnant bellies heavily weigh down on your cock during sex and that is likewise very hot
>>8639671I don't trust them, they managed to make noob worse rather than better with their finetune ( https://civitai.com/models/1095750/neta-noob?modelVersionId=1230787 if you never saw it )
They can't even improve a proven base model like noob so how could they possibly make a good full finetune of lumina?
>>8639758whatever helps you sleep at night, fetus fucker.
>>8639761isn't netaart what illustrious was based on?
file
md5: 1a6ce3e8770b757fd5c2ddc9b3fd2590
๐
>>8639678downloaded it locally and it does in fact do nsfw
>>8639764no
this is what illustrious was built upon:
https://civitai.com/models/162577?modelVersionId=183131
>>8639761tbqh if they manage to make a decent model it may show that lumina is actually just a good foundation to work from
>>8639763The cervix separates the vaginal canal from the womb anon. I can't speak for all pregfags but I'm not even thinking about the fetus when I goon to preggers
>>8639671The samples I saw on /g/ were all extremely melty, blurry and painterly at the same time.
>>8639783>melty, blurry and painterlyquasarcake, tianliang massivedonghuehue, (((ask (askme)))), ningen mame
>>8639787are you having a stroke
>>8639790forgive him he saw a prompt and output artists
>>8639790just naming the origin of melty, blurry and painterly
forgot ciloranbro, another favorite of the neta crowds
>>8639671>>>/g/105672646>>>/g/105691786>>>/g/105693084>>>/g/105693132>>>/g/105693167>>>/g/105693273>>>/g/105693286>>>/g/105693724
>>8639671>>>/vt/101579335>>>/vt/101579712>>8639803seems promising but underbaked
>>8639803/hdg/ lost again...
>>8639803this just looks like sd1.5 tier shit
>>8639789>network dim 16 alpha 32>conv dim 32 alpha 16is this intentional lol
>>8639827That config is just weird. Results are also pretty bad, especially the garbled hands thanks to the high te learning rate.
>>8639817Pony V7 will have tough competition when it releases.
>>8639789>>8639755https://litter.catbox.moe/4ddnu23vtaa6iwh3.toml
In other words, settings barely matter. Make sure you tag everything, get various view angles and zoom levels, and crop/mask any concepts that would be difficult to generate with your base model.
>>8639839I mean more or less, 90% of the testing I did was just optimizations, the DIM/Alpha and LR is pretty standard
Well beside the regular Adam stuff, that was enlightening in how people are using quantized garbage without even thinking about it just because some shmuck said so in 2023.
>>8639844Are you the guy that tested the settings of vanilla ez skripts one by one ceteris paribus? not to be rude, but that's just wasted time and hardly allows for any conclusions to be drawn.
>>8639846I don't really give a shit about conclusions, I was doing this for a setup to use and I have used it for 40 loras lol.
>>8639844I feel like most of them just make training go faster, don't seem to affect the end result much.
>>8639846Never heard of him, but sounds better than just guessing or parrotting opinions. Even if some of the conclusions only hold true for his testing dataset.
>>8639839Any advice on tagging?
The model can do porn. Doesn't look great though.
>>8639612What are your settings and tag number? I think my dataset is pretty good, it's more recent stuff with the plain ones left out, since I wanted the intricate shading.
>>8639863I mean, what tools did you use? Any NNs? I cannot fathom going over, like, couple of hundred pictures and adding varieties of "1girl, standing"
>>8639864Any wd-*-large-tagger-v3
>>8639862I just used the tags from the boorus. I had 182 images. Here's my toml
https://files.catbox.moe/49zz21.toml
>>8639864NTA but just use EVA 0.35 and adjust the tag min threshold according to your dataset
I barely even clean them up nowadays, just fast pass for shit like alternate x and such but that's not even necessary
>>8639860If you wanna do it manually, pretend like you're trying to generate the dataset image and the captions are your prompt; just leave out quality tags. That's what I did anyway. If you wanna use an autotagger, I recommend using a lower confidence threshold. It'll give you a bunch of hallucinated tags and also variants like "skirt, red skirt", but they're easy to clean up after and it's less likely to miss anything important this way.
Wouldn't use either approach with a larger dataset cause it's too much work, but the fewer images you have the more important accurate tags are. You can also do both and combine the result, then remove duplicates.
>>8639863wrong thread
>>8639440
>>8639839so i think my issue this whole time has been the fact that i was training my lora atop a shitmix. training on the base model instead has already netted me better results.
>>8639894Especially if you then use it on a mix. Because any issues you baked in will be stabilized by the mix the same way it stabilizes the base model. At a slight cost of accuracy, just like how mixes tend to have a bit weaker character/style knowledge.
>>8639894yeah that will do it, we tell people not to train on mixes for a reason
>>8639665 (OP)Oh hey someone actually baked, very nice
>>8639899you're welcome, you can repay me with 500 trap (formerly otoko_no_ko) gens
>>8639900What if I free you from your painful existence instead
>>8639905i will also take payment in straight shota paizuri (formerly paizuri) gens as well
>>8639907onee-shota is based I'll give you that but I don't quite like the idea of doing paizuri gens around that
>>8639861yeah the anatomy looks weird on this model. i'm having fun with it but it needs more time in the oven. feels better then chroma but that might just be because it gens way quicker.
>>8639908i like it because i like my shotas with small penis and the boobs can fully envelop the dick but shotas with small penises fucking women just looks silly
What's with this thread? Why is it separate from hdg?
Did I lose some kind of lore
>>8639912If you need to ask you don't need to worry about it
>>8639912we decided to keep /hdg/ as the community toilet and to designate this place for actual discussions like lora baking, new models/archs, issues with genning, is X addon snakeoil or not, etc.
>>8639912no reason at this point, they're about the same
originally a desperate attempt as escaping from one spammer
some people believe that shitposters will stay only in the original general, so they'll keep baking this
>>8639912/hdg/ went to shit a long time ago and each time I check back on it, it still looks like a bunch of garbage. So that this thread exists for me to actually discuss image gen is nice.
>>8639914Aight
>>8639915That's pretty cool
I've been wanting to learn how to bake loras for a while now, but I'm not sure my hardware can take it and it seems like such a pain to set up
>>8639916Count me in on the hopium
>>8639918Have you checked today?
>>8639919Can you gen using illust-based models? Do you have an nvidia gpu? Do you have at least 8gb of vram? Are you at least on the same level of retardation or less then I am? Then download https://github.com/derrian-distro/LoRA_Easy_Training_Scripts and you can train loras using this toml
>>8639866 or one of the many others linked.
>>8639925>Can you gen using illust-based models? Do you have an nvidia gpu? Do you have at least 8gb of vram?Yes, I run a 3050, it's exactly 8gb of vram
>Are you at least on the same level of retardation or less then I am?I guess we'll find out, thanks anon
>>8639927I'm also on a 3050 so that toml will definitely work for you.
>>8639900>formerly otoko_no_koBased
>>8639905>free you from your painful existenceSpoken like an ugly_bastard (formerly ugly_man)
>>8639491>291hMy beloved. Still my daily but I've switched out 102d custom for 102d colorized. Not sure what that nigga did but I'm getting cohesive backgrounds coveted by those of Indian descent and not impacting artist integrity. An easy upgrade.
>>8639915so /hgg/ is for technical support and /hdg/ is for gens and shitposting?
>>8639944>/hdg/lmao
anyway go back
>>8639944you can also post gens here
>>8639929Cool, any noteworthy tutorials I should know of before going sperg mode?
>>8639938Is it (formerly) because the other tags work better?
I've ran some tests before and otoko_no_ko and trap do give different results, but different enough in intent that they're both notable
>>8639923No. I assume it ebbs and flows anyway even if today is a good day.
>>8639966uhh i think people liked this one but it's old as shit and for sd 1.5
https://civitai.com/articles/138/making-a-lora-is-like-baking-a-cake
just don't crop images unless you have a good reason, remember to make sure your images are tagged and use this to tag them if not https://github.com/jhc13/taggui. otherwise it should be plug and play just replace all the names
also you might need to use this fork of easy training scripts
https://github.com/67372a/LoRA_Easy_Training_Scripts
im not sure if the original has updated with all the settings i use now that i think about it
also no they just changed the tag names on danbooru recently
the old tags will still work better because no models have updated to use them yet
I've seen people talking about removing character tags from their datasets, why do you do that?
>>8639966>Is it (formerly) because the other tags work better?I'm just meme'ing but as I understand, the trap and ugly bastard changes were recent so none of the models would be trained on them. At least trap. ugly bastard is a tag on e621 so that would work but from that dataset.
>>8639973i don't want them in there
Can anyone help me out with this character LoRA I'm trying to bake? I've been trying to successfully bake a character LoRA of this character for the past 3 weeks for noobxl and I've been facing lots of issues with her prosthetic arm (and I've tried almost everything to solve them), her prosthetic arm appears on the wrong side of where it's supposed to be around 40% of the gens and sometimes it even gives me double prosthetics, specially when I prompt her from behind or from side, I've tried different configs with different LRs and Dim sizes, and yes I have flip_aug off. I've made sure to get enough pics from different angles to solve this issue but I'm still struggling with it. Is there a way to solve these issues for characters with asymmetric features like this one?
>>8639984What's your batch size
>>8639984share the config you had best results with so far
how many steps are yall training character loras for
>>8639985I've tried from just using 1 alone to 4 with grad accum 4 (with adjusted LRs ofc) and I've seen no huge difference tbqh
>>8639987I think this one was the "best" out of them all
https://files.catbox.moe/72zp6j.toml
I use keep_tokens for a trigger word with her name
I bake for several epoches and try lots of them to try to pick the best one
Also I've noticed that if I increase the LR a bit it starts to bake some sort of style in, how can I avoid the style bias?
>>8639993My nigga
>>8640002I could always try baking it myself out of curiosity, gimme that dataset. Probably tomorrow though.
Worth nothing that I think using shuffling captions fucked my shit up when training unusual characters. Nothing else that could cause it, I don't think. Rin was a bit harder to bake and Emi just refused to.
>>8640002>Is there a way to solve these issues for characters with asymmetric features like this one?This can be pretty hard for the AI, yeah. I remember baking a pretty decent Baiken LORA back in the SD1.5 days, it came out well enough but I'd get "mirroring" issues all the time with her eyepatch and missing arm. It can help if you avoid stuff like Flip Augment, but it looks like the toml you've linked doesn't do that. That said, this is just a measure that helps prevent the AI from fucking up, I don't know anything that causes positive results, aside from the usual advice of "Find a bunch of high quality images that don't fuck up the robot arm placement".
As for baking the style in, that can be inevitable as well with high LR/epochs. The AI is rewarded when it converges towards something in your dataset, and eventually this means the AI will optimize to match the style of your dataset, if there is a predominant one. I imagine you relied mainly on anime screenshots, which is fine, but try to mix in more fan art if you want to avoid style being baked in.
Does someone have the script for right click to view metadata? I wrecked my computer and had to get a new one
>>8639984>>8640027this is why I hate sdxl with all of me.
i really sure wonder what this mysterious script that allows you to view metadata everyone talking about is
it certainly can't be in the op as i always check the op as i am smart
woah watch out we have a man of EDUCATION in the thread
>>8639970Sorry, do I need to install triton for that version of easy training scripts? It says "Make sure to follow the steps at https://github.com/woct0rdho/triton-windows?tab=readme-ov-file#install-from-wheel for Windows" but the link just goes to the main git page
Thanks for all the resources anon this thread is awesome
>>8640044i dont think you need to install triton it just speeds things up but if you do you just run "pip install triton" while inside the venv
>>8640044triton doesnt work on windows unless you manually install it, dont bother. safe to ignore
>>8639984This is a common problem even with existing characters. Yang has almost 3K images and noob still often gets her arm wrong. Same with many other asymmetrical features. I don't think there's a way to solve it while we're stuck with SDXL.
invest in neta lumina if you want to be done with a soon to be two year old architecture
>>8640004>Rin was a bit harder to bake and Emi just refused to.nta but who? im curious
>>8640054guessing thats the katawa shoujo anon
>>8640058oh
how shit sdxl could it be that it couldn't learn how to do prosthetic legs like what the actual fuck
this garbage is turning 2 years old next month, HOW ARE WE STILL USING THIS SHIT
>>8640077Is there any actual competition?
I can't believe I'm gonna say this, but, kurumuz, I kneel. I feel like I'm a caveman with sdxl.
Speaking of SDXL problems, I baked a character LORA recently and for whatever reason, the hair has a bias towards having a purple tint. I can't figure out why that is, none of the images in the dataset have purple-tinted hair even a little, it's a character with dark-brown/black hair depending on the lighting. I'm even prompting for dark hair. If I put purple_hair in the negatives it's fine but that's still so odd.
I also have a hard time getting characters with atypical proportions to work just because of finetune bias, but I guess that's an issue with Illustrious. I dunno why a shortstack character LORA+short stack in the prompt wouldn't work though.
>character has almost no fanart or official art outside of an anime
oh boy style bias here i come
>>8640105Perhaps you could use ChatGPT/Gemini to make some pics? Never tried those for such a task. Like give them a character sheet or the closest you can get to one, and then ask them to generate this or that image with this or that art style. You could also get another image and ask them to replace the character in it with the one you want. And you can work with it in some back and forth to try and get certain details more accurate if it gets them wrong which I assume will likely happen.
>>8640110Gemini wouldn't be great for that because it can't really fix things and repeated edits will do nothing but degrade the image into orange slop so it has to be right the first time
Probably alright for simple designs but not anything complex
>>8640110I've used some synthetic data before to try to reinforce characters that don't have many "full body" examples, or are maybe only viewed from a specific angle, and I'm not convinced it's a good idea. I could see an argument for taking a style-biased LORA, applying an artist LORA to an output, then feeding that kind of data into another LORA but avoiding style biases can be done a lot easier just by not overcooking your initial LORA. It's harder if a majority of the art comes from a single source but not impossible.
>>8639866Sorry for being vague, I meant like the actual number of tags per image. Ie 30, 50, etc. I just started baking but I noticed that when I went truly autistic and had 50+ tags per image (self curated, it was torture lol) thestyle lora somehow got weaker.
>>8640169thats not something that should actually matter unless your training settings are fucked up
Man... at this point I'm pretty sure my gpu is fucked in some weird way.
Like literally none of the configs from the last year works, producing either warped and mingled garbage on epoch1, or showing literally zero style transfer.
>>8640154That process is also mostly unnecessary on noob since it knows so many artists and even meta art tags, you can just put than in your captions.
Not like either will get rid of the bias, but it will make it much easier to override the style with another when genning.
>>8640284Would have to be very weird. Like with generation, as long as you manage to get an output and not a black screen or error it should always be the same output, regardless of hardware.
>>8640288Yeah, generation works just fine, but any training ends with garbage. Even tried to install a fucking Linux - the result is still the same.
Like, every config from these threads: 50 different datasets starting from 20 booru-tagged images and ending with 300 manually-tagged fanbox rewards. The best I've got so far - a weak style transfer with melting eyes and fucked fingers.
>>8640154>synthetic data to reinforce characters that don't have a specific angle>not convinced it's a good ideaWhat's the downside? I have yet to try this, but my girl has no views from the back. So it's either letting the AI guess what she's supposed to look like with every gen, or have it guess once then refine the result with inpainting and lock it in.
>>8640077anon
i said it was the shuffling captions, it worked in my older setup funnily enough
or i fucked up the newer dataset
>>8640305IMO the biggest theoretical issue is that you could introduce a lot of noise into your LORA that you don't get with real artwork. Remember, a model just de-noises random pixels into something that looks like a created image based on text input. This also means that noise doesn't totally go away, it just doesn't seem there at the end of a generation to the naked eye. Putting noisy data inside a training dataset will cause a feedback loop of the model trying to converge on more noise, which results in lower image quality. I think it's probably okay to try, especially with higher effort inpainted stuff, and/or if your dataset has a lot of non-synthetic content, but don't rely on it. For what it's worth, when I experimented with it I had a similar issue and it didn't seem to improve the results. It wasn't terrible or anything, it just didn't seem to move the needle to it ended up being wasted effort.
>>8640313DESU baed on my experience with other weird datasets and images I'd say there's a good chance this is leftover advice from 1.5.
SDXL seems really resilient to baking in garbage of any kind compared to 1.5, for good and bad.
>>8640310actually i ought to test this, this dataset was weirdly shit to bake and i do think it's shuffling, there's not much else that could be breaking it
>>8640317The one time I baked a fully AI dataset it was very prone to missing limbs, and none of the dataset images had broken anatomy. But not sure how much one or two pics would influence an otherwise normal dataset.
hm no still getting shit results
i guess this setup is good for everything but cripples lol
>>8640352>no shuffling gets me hands and legs>shuffling gets me no hands and no legstime to do a half shuffle lol
why do we shuffle anyway?
>>8640373Well I tested it and it was just better overall. In this particular case not shuffling baked in major style bias compared to shuffling.
>>8639938>102d colorizedGyad damn that's some heavy bias for empty prompts
>>8640010I've liked kotteri for a while
>>8640375Hit me with some style grids, autismoanon.
>>8640399this was enough for me to immediately delete it kek
>>8640375my bias is 1girl, large breasts
>>8640404Based heteroanon.
>>8640375All these shitmixes are like that, how people can stomach using them is beyond me.
>>8640410how do you use noob vpred 1.0?
Impressive. Very nice. Now let's see no snake-oil's vpred 1.0 anon's gen.
thiso shitto again
there are 13 images here nobody's using noob or shitmixes or anything
>>8640417of course nobody posts in this shithole. but we still gen
Does random clipping really improve generalization?
>>8640418i don't trust you
uh.. base vpred 1.0 bros? Our silence is deafening..
actually these generals are great for gen advice
just do the opposite of what is said because it's always some weird form of trolling
>>8640426didn't we already address that a week ago
>>8640430our response, /hdg/?
What checkpoints do you recommend for style accuracy anons?
Not just anime, but overall, even for more oil-painting looking styles
>>8640450"style accuracy" is just danbooru knowledge, in which case you start with v-pred 1.0. Then figure out if it looks better with a bit of slop merged in, at the cost of diluted knowledge. Slop merge ranking:
>1.0 2d custom>ฮฃฮฮ>WAI-NSFWIf you have a specific style in mind that doesn't have booru images, use a lora. Above still applies.
>>8640450What anon
>>8640452 meant was
>291h>102d customYou can stop there and switch as needed.
>>8640456I meant to give him the full spectrum
>>8640452>>8640456Sorry, I'm a bit of an illiterate retard since I've learned everything by trial and error rather than actually reading
v-pred 1.0 is noobAI right? Is there any difference between noob and illustrious, or are they the same structure with just different knowledge?
Also where do I find those models, except for WAI NSFW which I know is on civitai
>>8640467>v-pred 1.0 is noobAIYes, that is the final one.
>any difference between noob and illustriousNoobAI is a finetune split off from illustrious 0.1. Trained a lot more on Danbooru and also added e621 knowledge.
The illustrious team continued training their model to 1.0, 2.0 and now 3.5. but they don't seem to be nearly as good as noob.
>where do I find thoseAll are on Civitai except for illustrious 3.5 which is paywalled, elsewhere
>>8640467https://civitai.com/models/1201815?modelVersionId=1491533
https://civitai.com/models/1301670/291h
>>8640450you're not really getting many good style outputs even out of shitmixes
artists are fucked up by vpeed and need to get fixed and all of the alternate styles are pretty weak
>>8640470>>8640473Pretty cool, thanks
On this topic, what are some good ways to test different checkpoints for quality?
I have my own of course but like I said, I've learned how to use stable diffusion on my own by trial and error so I'm curious as to what you guys will say
>>8639984>>8640051You're not wrong but 40% seems high. I imagine those 3k images are bad because the artist drew the things on the wrong side. Moles might never work, but something as big as the prosthetic arm on the correct side on a character lora should be as easy as having it on the correct side in 100% of the images. SDXL might still mess up but it should be closer to 10% or even 20% mistakes not 40.
>>8639612>>8639658What are your settings?
i feel like baking, but...video games........
i feel like genning, but.../hgg/........
how is noob vpred 1.0 unironically the top recommendation here, half the artists feel like the model is on the verge of exploding
I only gen with vpred 1.0, no snake oils at all. I cannot believe people can tolerate shitmixes. No, I will not post a gen so you can gauge the quality of my gen.
>>8640562Everything this place states if wrong. If this place states you should use 1.0 then you should be looking for a shit mix.
>>8640562it isn't
it's like two retards between the two generals
>>8640562name a model with better backed-in styles, and without 1girl, big boobs, standing bias
>>8640572>1girl, big boobs, standing biasdoesn't nai v3 do that if you just gen with quality tags as well
>>8640562Stop using Euler a Simple 5 CFG, retard.
>>8640562its not
>tell me what isnyo
>>8640562I also use it the most, but either with my own loras or a stable style mix. So it's not really relevant to people just trying it out.
>>8640532NTA but I just do a bunch of gens and get the natural feel of the model lol
Not the most precise method
>>8640532what is that question lol
you gen
you look at it
you choose what you like
>shitmix vs base model ""discussion""
>again
jesus, just use whatever works for (You)
but there's one gabillion models. i need to be spoonfed the bestest golden model i can put on my hard drive
>>8640532>ways to test different checkpointsRead up on how to generate XY plots, in whatever UI you're using.
>>8640562I don't think it's a drastic difference. Anyway the advice I took from earlier threads is to use a weak 0.2 or so strength lora on top of v-pred 1.0, one that somewhat matches the style you're going for. Which feels about the same as using a shitmix for stabilization except you get to choose the style bias, rather than it always being Nyalia-adjacent with semi-3D and shiny skin.
>>8640620itd be outdated within 3 weeks. and you know nobody maintains rentries
>>8640620yeah except the advice would be "whatever works for you" because people used 15 different models with 20 different strong opinions about them
>>8640621niimi my beloved
>>8640622>itd be outdated within 3 weeks.huh?
>>8640621This, shitmixes are OK if you are using some limited online generator or you are extremely lazy and just want to look at some boobs in generic shiny AI style.
>>8640621>no r3mixit's my favorite shitmix
>>8639983just prompt for ixy
>>8640638t bh I have enough trouble telling apart these few I have saved
substitute any variant you like
thanks
need to troll /hgg/ into merging back with the main general
>>8640610Not super practical when your setup takes 2~3 mins to change models
>>8640621Yeah this seems to be what I'm looking for, thanks
>>8640659He's kinda cute actually
>>8640673>2~3 mins to change modelsdo reforge fags really LOL
>>8640679I don't have an ssd grrrr
file
md5: 7000445c06eca0d2aaa115438ca18e75
๐
>>8640679idk what that anon setup is but it takes around 1 min for me to change models on reforge
>>8640624Really don't see her much in AI threads
>>8640681>>8640679it's a disk speed issue, UI might add a second to it at most
needs to read 6.7GB from the drive which takes forever on HDD
>>8640681I meant hardware setup
>>8640683I really gotta stop being a coin pinching jew and buy a dedicated ssd
But where I live a 1TB one unironically costs half of my salary and it makes me hesitant even when I have the money
>>8640682I did some gens of her on the NAI v3 days, maybe I should remade some of those
>>8640685it's not exactly penny pinching if a 1tb ssd costs half your salary, man
forge classic has really good model loading time
too bad it's all fucked up for genning
>>8640693Oh I think he actually fixed that stuff up
Although I still can't get parity because of sampler availability, I gotta test between the same samplers.
>>8640717Okay, it's still just a BIT shitter on classic. although not as bad as before. I wonder if that's fixable with settings fuckery.
>>8639737Pregnancy has to be one of the most primal kinks you could have. Pregnancy = Fertility = Can pass on your genes for you. Every ancient fertility statue is pregnant women for a reason.
Does genning pregnant waifus without genning pov, cum in pussy beforehand make you a cuck?
>>8640412You use a very low CFG (2-3), and recommended Euler A sampler for 25+ steps. Then it's all about finding a nice artist mix. That's literally it, 0 snakeoil required no matter what everyone else is saying.
Anyone got some tutorials for creating my own comfyui workflow? Using reforge atm but it not being maintained is worrying me.
>>8640757>but it not being maintained is worrying me.how so
>>8640757stick to default nodes. dont fall for the efficiency nodes meme
Okay, removing the argument for forcing ZSR noise has gotten me almost 1:1 to reforge but still just a tiny bit shittier. Weird.
>>8640757What do you think you need a tutorial for?
>>8640741That would be impregnating (goated), not fucking pregnant women
The latter just reeks of low impulse control (on a porn board? no way)
>>8640717really? which one is best now? i'm still on a fucked up version of reforge because updating broke everything
>>8640741How can you tell a woman is capable of carrying your child? Only if she is pregnant will that be true. I don't know what you mean by "fucking a pregnant woman", the idea of her being fertile is arousing, has nothing to do with fucking her at that moment.
>>8640772Counterargument: https://files.catbox.moe/2jzhxv.mp4
>>8640804>>8640772Disregarding the raw allure of the pregnant belly: satiating a womans pregnancy-borne horniness is also very hot
>>8640717You don't NEED more than euler.
pregnancy is /d/ btw go back
>just realized reforge stopped development
wtf do i use now
>he was using a commit younger than 2 years
NGMI!
>>8640809Dumbest and gayest shit the mods ever pulled.
I get hyperpreg but pregnancy of reasonable attractive proportions is very /h/
It's not even all the mods either. It's 50/50 on whether a preg thread survives.
this place was better when people posted nai gens
>>8640820Null's fantia paywalled behind 300 yen
Sankaku has the non-preg versions of these animations and pics.
Also theres this: https://files.catbox.moe/x10t59.mp4 (he didn't make a preg version of it)
>>8640818>Dumbest and gayest shit the mods ever pulledNo that was making tentacle porn /d/.
>>8640837wasn't that like 2003
>>8640837That's at least halfway understandable. Even though I agree with you.
>>8640847>That's at least halfway understandableHow?
>>8640849Nta but isn't that basically bestiality?
>>8640855>tentacle, the quintessential monster that fucks anime school girls, is somehow not hentaiThere's a reason hentai is considered "perverted." If tentacle is not /h/ then /h/ should have been named something else. All that is before we start talking about orcs.
>>8640855So is blacked but yet you see it everywhere
>>8640855Thats why I said "halfway understandable"
>>8640862lel. tentacles & soft-preg switching places to /d/ with NTR and Blacked would be kino
>>8640871>halfway understandablebut you're wrong?
>>8640876Well ok anon, I'm listening.
>>8640855bestiality is hentai
it's kinda fucking retarded the monkey fucks schoolgirl while tentacle prods her ass genre is representated by generic 1girl sex here
>>8640871>tentacles & soft-preg switching places to /d/ with NTR and Blacked would be kinoBased
>>8640871>>8640890M4GA (make 4chan great again)
>>8640860>tentacle is somehow not hentainobody said it isn't
/d/ is "hentai/alternative", a subset of /h/ for less normie fetishes
>>8640879Eh maybe you're right and it's just total bullshit tentacles aren't /h/
I think /h/ general definition is: Vanilla and vanilla adjacent fetishes conceivably performed in our own world. More or less.
Style of girl
Girl doing X act
Sex in x situation
/d/ is dickgirls, bizarre shit, and stuff that can't be emulated in this world, plus bestiality cause that qualifies as bizarre.
This is why tentacles are "halfway" understandable to be a /d/ thing
>>8640899we could just, you know, move to, you know, that other blue looking imageboard website
>>8640902bestiality isn't /d/, either because of laws or because of its adjacency to furry
file
md5: 6354a433788a7e6a082809e462ebd341
๐
>>8639665 (OP)My pic was OP? Nice!
>>8640908is that butakoma?
>>8640907>its adjacency to furryThat's precisely why it IS /d/
>>8640901>normieHentai is not for normies anon. Anime isn't even for normies. Normies would, and still do, make fun of you for fapping to cartoons.
>nobody said it isn't"kinda understandable" means that it kinda isn't /h/ and kinda should be /d/
>>8640902>can't be emulated in this worldI thought monster girls were allowed here?
>>8640911His shading is different, both skin and hair. Can't rely on that tho as so many people use WAI now which would override it.
>>8640913anon furries are /trash/, they're not allowed on /d/ and neither is bestiality
file
md5: 373db3bc2cd9813a70f1727c5a98fbe2
๐
>>8640911Yes!
https://files.catbox.moe/wnw612.png
>>8640916Eh, edgecases.
Those are A. female so the bar for bizarre is higher and B. they're lean-human-humanoid so it's cool.
>>8640916Everything's allowed here, as long as it doesn't provoke people into reporting you. Monstergirls are fine as long as they just have ears tails or horns. The rest you have to feel for yourself, it gets risky if they have goat hooves or fur on hands/feet, or a different torso like a spider/lamia. Centaurs are an instaban even though spiders are alright. etc.
>>8640921Oh yeah, I actually forgot it's loli AND bestiality that are banned from /d/.
Because of countrys' laws. I forgot that bestiality's illegality influences /d/'s decision on the matter.
>>8640922mind sharing the lora?
>>8640924>femaleOrcs, kobolds, that monster from genshin everyone likes, goblins, demons, etc. are allowed here and no one cares, but I guess anything I bring up is also an edge case? Mods are stupid and janitors are even worse, anon. They don't care about board culture.
>>8640929>Everything's allowed here, as long as it doesn't provoke people into reporting youThis is really what my post was leading up to.
makes you wonder why it's spelled "best" instead of "beast"
Anons is this a good model?
https://civitai.com/models/934492
I've been using it for a long time and it beats NTRmix and WAI NSFW imo
The best model is the one that works for you
>>8640936Quit trying to get me to exhaustively cover all cases. You know the GENERAL vibe of /h/ and that vibe is what commands the rules.
Yes it's to a certain amount arbitrary but my challenge to you is thid: why shouldn't it be?
file
md5: 0b7a16792fcb9d93905c720a0d022455
๐
>>8640932Needs his name as an activation tag since I was trying to overwrite the original with his newer stuff. Weight of the tag is the better way to weight the lora.
https://pixeldrain com/u/rCNYpUpM
>>8640948I'm guessing it just happens to fit whatever style(s) you use better.
>>8640963Well, I like varying styles a lot
From 2D to oil-painting
>>8640940GPT says:
The correct spelling is bestiality, not beastiality, due to its Latin origin.
Here's why:
Bestiality comes from the Latin word bestia, meaning โbeast.โ
From bestia we get bestial, meaning "like a beast" or "brutal," and bestiality, meaning "behavior or acts like that of a beast" โ specifically, sexual relations between a human and an animal.
So why not "beastiality"?
Beast is the English word that also comes from bestia, but beastial is a common misspelling and mispronunciation of bestial.
People often insert the extra "a" because "beast" feels more familiar or logical in modern English, but it's not correct by etymological standards.
In short: it's bestiality because it's derived directly from Latin bestia, not the English word beast.
Let me know if you want a breakdown of similar words where spelling diverges from expectation due to Latin roots.
>>8640953>GENERAL vibe of /h/Yeah I'm saying the "vibe" is stupid since tentacle should be /h/. There aren't any good arguments for why not other than "I'm just not feeling it" as you've now pointed out. Hentai as a thing is inseparable from tentacle as far as I'm concerned and is as important to the vibe as magical girls, sailor suits, and plugsuits. That's the correct "vibe." It's the quintessential hentai, regardless of what retarded normies think.
>why shouldn't it be?What do you mean? Why shouldn't what be?
Bestialidad se escucha re cojudo comparado con zoofilia
>>8640981that's disappointing
>>8640904>move to a dead wastelandhmm... nyo
>>8640983I'm saying the vibe is right except for tentacle, the sole exception.
>There aren't any good arguments for why not other than "I'm just not feeling it"And? That is THE argument concerning fetishistic taste. Don't be an autist.
Don't be so hung up on the fact that the lines between things are fuzzy rather than hard.
>>8640997it wouldn't be a wasteland if you faggots did your part
>>8641009But I like the rules here
I'm gonna try baking fully on SPNATI sprites and seeing what happens
>posts 281, images 21
Come on faggots this is why /b/'s threads are so much better even if they are a little too fast.
Just. Bake.
Just. Mindlessly. Fucking. Bake.
why are zoomers like this
>please spam garbage slop
im good chud
>>8641030>said anon in the ai threadCheck
>acting like the things that ever ARE posted here are some kind of long baked curated masterpiecesand Mate
>>8641025>/d/ddg>364 / 163>/d/dddg/152 / 90
>/trash/slop504 / 226
>/trash/sdg/>397 / 170>/b/degen/>150 image limit in the blink of an eye constantlyYep. What is fucking wrong with /h/?
just dump your slop into /hdg/
what's the problem?
>>86410521) They post cockshit
2) People here are not posting because every gen in the past 2 years would get you shit
>>8641015Would you believe me if I said it was completely style fried kek
Now I want to rebake it with "official style" tagged
But this is mostly a prelude to me trying to make a fully ai dataset to test if it really is bad to use on modern models
>>8640957nta but thanks
it mixes well too
>>8641054imo it's worse getting no response than negative ones, like why did I even bother
more so if it's an image genned for /h/ specifically
>>8641082do you guys really
>>8641070>fully ai datasetIt isn't and I've been doing it since pony.
>>8641087used to, yeah
more than one thread per day, you couldn't get that kind of activity without people sitting in the thread for hours at a time trying shit out or to riff off a gen someone else made, continuing either with the character or the concept
>>8641070You can do the thing. Train style lora, merge into model, train character on the merge.
>>8641093I have no idea what are you talking about broski but I'll post a fail gen of mine I can't get bother to fix at all just for the sake of it
>>8641099Why don't people do this more often? Is there really no drawback?
>>8641103other than her foot-hand it's alright
I would inpaint rather than call it a failure
>removed a few images
>removed some tags
>added official style
>it actually helped the style burn in massively
kek
definitely not perfect but i could probably gen a proper dataset from this
>>8641108The drawback is you have to train two loras. Prepare two datasets, etc.
And it's rarely useful on noob since it has enough artist knowledge, you can just tag the style instead.
>>8641110The failure is me for not seeing that part way earlier into the inpaint process, at that time I thought I was done with it but then I spot that aberration and lost any will to fix it
>>8641119? Not really? You'd just take screencaps from a vidya of random NPCs for the style lora then pics of your favorite slut and you're done no? My isn't everyone else doing this? You'd think the style burning issue would be solved by now but it doesn't seem to work that well in practice.
>>8641125>>>>>The drawback is you have to train two loras
291h really is a beast
Any consensus on the sampling method, scheduler, steps and CFG?
>>8641125It works pretty well. Will never get rid of all style bias, but makes it easy enough to push the lora into working with other styles. I've done it a few times on Pony but didn't have a reason on noob.
>>8641138Did you have to train the resulting character lora for more or less than usual? When I did it on pony it didn't learn very much.
>>8641140I don't remember. 3K steps on that ancient AdamW config, according to metadata. But I may have saved an earlier epoch.
>>8641137>consensusAHAHAHA
>>8641151Alright I guess it's all worth another try.
>>8641137>291h really is a beastat doing what exactly
>>8641166at having a slightly different style bias
We've entered the era when even the shitmixes got stale
Nothing new interesting, based64 situation again
>>8641170based64 wasn't even good it was just first and made by an anon
>>8641182first to what lmao
it came out laaaaate into the aom lifespan
>>8639737This man is requesting this multiple times in other AI threads.
>>8641183first to be a 1.5 model that actively tried to be style-neutral and a good base to use loras instead of having the generic aom style
>>8641170based 65 mix was the only good one
VGH I am finally back home.
>>8641177It was img2img, style variant of an older gen. Same prompt on a raw gen gives this https://litter.catbox.moe/7108xqdb8szbvrd0.png
Great Magami lora was from a /hdg/ anon, about two months ago? Check the archives.
>>8641184Different boards have different cultures on requesting it appears. /d/ and /b/ have been helpful and informative. /h/ and /trash/ a little less so.
I guess it's all in what anons you find. I probably won't find success with the overview approach and will probably have to take a one-pokegirl-at-atimke one.
Fingers crossed I hit the jackpost in a baker.
file
md5: 80e83a1ad3d0d615097d227c88ab8bbb
๐
I have learned so much this week:
I don't know what I'm doing baking.
I don't know if my set-up is even working correctly.
I don't even know if the model settings I'm reading on Civitai are current or not.
Is it always like this for you bros?
>>8641213If I am getting what I want. everything I do is correct
The blackbox slopmachine is all about RNG anyway so don't think too hard on that
>>8641213There are too many knobs to adjust, so I just cope with "settings don't matter" and bake on the same one from 2 years ago.
Did you get any decent loras out of it? I would like to have the one that produced your pic.
>>8641213nyo because i've spend exponentially longer testing stuff for fun than genning
>>8641224No lora with that one. Also, no, I tried baking with triggers and I got some strong dogshit. The only okay style I got was with anon's toml here
>>8639866
I don't know where to ask this, but what about voice generation? Like text to speech. Anyone here know the best way to do it?
>>8641311Local? Don't even bother.
t. bothered
>>8641311there's next to no discussion of it because no one locally can get even close to the results of elevenlabs and google
>>8641311generate a compelling image and ill fill you in on the sorry state of TTS
>>8641314>no one locally can get even close to the results of elevenlabschatterbox?
sesame?
>>8641317i tried chatterbox its good but still sounds clearly ai in non-cherrypicked samples
>>8641314>>8641313what about non locale? Anything free that doesn't suck too much?
>>8641323>anything freeno
>>8641314how is image gen the thing that require the less amount of compute
doesnt make sense to me
>>8641327i don't think it's really compute as much as just corporations not yet getting into the stuff
there are some tts programs that can get you minutes of okayish audio in 2 actual seconds
>artist giving me old ass style interiors
man i need to go a heritage park again
>>8641327>>8641328Picture an apple in your head.
Now try to say the world apple aloud.
Now try to describe what that apple looks like in words.
The difference in effort is the difference in compute.
lil bro just making shit up to sound smart
Please give me rentry or something that had artstyles list of Nai 4.5 . With example gens
>>8641333thats the dumbest comparison possible
replace "picture an apple" by "draw an apple" and you get a more realistic one and its still retarded
The problem with audio isn't compute, it's data.
>>8641327its not compute related, it's technology related. the only people making strides in audiogen are keeping all the discoveries to themselves.
>>8641343https://tagexplorer.github.io/#/artists
>>8641317sesame released 1b trash and kept the actually good 8b model to themselves along with the steps to actually replicate the website demo
>>8641365kek
open source always loses
In nai 4.5, how i can get 90s retro anime style like this ? https://civitai.com/models/1060869?modelVersionId=1245262
>>8641372just prompt it
>howbrowse danbooru
>>8641372No idea for NAI but if it goes by danbooru rules then anime coloring, anime screenshot, official art, on model, retro artstyle, 1990s \(style\), film grain, and whatever anime you have in mind (case in point, evangelion) should work
>>8641365>sesameYou wouldn't release your krabby patty secret recipe neither.
I've been trying to learn Forge, what's the best way to set up multiple characters in a scene? Even with the Forge Couple extension it mixes elements alot of the time and ignores other prompts.
what is the prompt for thighhighs that end right below the butt cheeks?
>>8641586Yeah I wish that had a tag a long time ago.
>>8641586don't know there is one, but it's easy to inpaint
>>8641137>291hAh. A fellow absolute nigga with an aversion to ai_gloss. I like you. DDPM / Beta or Euler SMEA Dy / Beta is all you need. If you're craving a bit of creativity, try AYS Custom but it has a few side effects.
>>8641213>Is it always like this for you bros?291h keeps teaching me that no matter how much I meme myself into thinking I found a model that does something better because of RNG, I always come crawling back to it after a few days.
>>8641675only wish it used ฮฃฮฮ instead of ลฤฐH
there was no need to include Better Days, looks sloppy and it's a base illu merge
>>8641688Better Days means Better Gens, anon. Shame the anon died before completing his magnum opus. 291h mogging all models while incomplete.
Anon-kun, there is a report function, you know?..
>>8641709/d/ is for futa, which is not noice
This finally baked somewhat competently
Still not as good or artifacted as I'd like it but eh
game cg https://files.catbox.moe/30mbse.safetensors
>>8641725they have a separate non-futa general now
though it is a bit slow
>>8641730About fucking time.
>>8641738since early February
What do you guys use for manual tagging? Looking for something minimal - like image preview, tags list, sorting, autocomplete.
>>8641752BooruDatasetTagManager
>>8639871>EVA 0.35What's that? Is there something newer than eva02?
Finding an artist with a large imagebase and internally consistent style for baking is my new fetish
>>86417580.35 confidence
>>8641761Huh, isn't that a bit too low for eva? I remember HF page saying about ~0.52
>>8641761Here https://kemono.su/patreon/user/19728842?o=50
Up to 11/2022, then he moved to AI slop.
Speaking of Better Days, is there a way to get its funky lighting on base vpred? The author made an lora extract but it kind of sucks.
>>8641766That's what I personally tested for, YMMV ig.
>>8641770I'm doing some japs from the cg dataset now, I really like hashimoto takashi
>>8641778I thought 400 danbooru pics was on the edge of working well enough in its own.
>>8641781Eh I already know that vpeed ruins artists even that are pretty coherent without a lora
>>8641752https://github.com/starik222/BooruDatasetTagManager/releases
>>8641726How did you do it?
>>8641808I just severely cut down on the dataset to choose only the stuff with the moeshit eyes, but it's still weak for making like westoidshit look animu
The specific artist ones are probably gonna be better, I got 4 which I liked from that dataset and rescraped them fully
>>8639900>repay me with 500 trap (formerly otoko_no_ko) gensbest i can do is 1 for now, but i tried making it high quality so it counts as much for at least 50 gens:
https://files.catbox.moe/cqz6qn.png
>>8641781direct picrel
noob is just fucking ass with contrast because of vpeed even with all the shitmixing
>>8641889i mean lmao
https://files.catbox.moe/ikxhq4.png
Why is noobai so weird
It's got great prompt adherence but it's weirdly unstable
Base style lacks detail
It requires a specific weird ass set of settings (euler + very low cfg)
No hires fix or inpainting settings I've tried look good
It feels like it spends 80% of the steps as a blob before actually making the image in the last few, which is the complete opposite of every other model I've tried before
It's so weird
>>8641953prompt an artist.
>>8641958Yes I've been using several
The main one is yuming li
Not saying it's bad, it's just very weird and not what I'm used to
>>8641953They tried to be edgy with vpred (which I think for anime is a wholly fucking retarded idea and was likely done for butiful background civit folk) but trying to bake it onto eps base and an eps tune with predictable results.
Also with mitsumi misato which is one of the best recognized artists on noob
Noob is just shit lol
>>8641981stop using base vpred retard
it sucks ass
>>8641978I have no idea what any of that means
Speak in layman's terms
>>8641983That's not base vpred retard, that's how shit noob vpeed is, even shitmixes can't fix it
>>8641985Vpred is a noise schedule that supposedly allows for deeper range of colors and better contrast, especially regarding night shots and such
But for anime it makes so sense because even night shots in weeb images tend to be evenly light and not chiaroscuro meme garbage
>>8641987Is that why it spends more than half of the gen as a dark blob?
I have mixed feelings with noob, on one hand it's danbooru+e621 knowledge and prompt adherence blows all the competition away, it's great at doing all sorts of styles which I love, and it seems to be good at taking all sorts of loras as well although I need to test it more
But on the other hand it seems to take some time getting used to and it's very restrictive regarding the settings you can use (particularly if you don't like euler like me)
>>8641997I think the main problem with it is as always, base was really underbaked and needed serious stabilization and people just bake garbage into it. Same as 1.5, same as Pony.
Vpred certainly doesn't help though, either it or the way Noob implemented it is just not really compatible with actual anime, it's for 3dpd stuff imho.
>>8641983>stop using base vpred retarderm i thought vrped 1.0 was SOTA
>>8642000that's a shitpost and has always been a shitpost
>>8641997Basically my experience so far has been
Cons:
>Restrictive settings>Shitty and inconsistent hires fix, adetailer, inpainting, prompt editing and color grading>Kinda shitty base style (non-issue with artists and loras)>Lower detail (backgrounds can be bland if you don't actively prompt them and it's INCREDIBLY inconsistent with smaller details)>Lower quality overall (somewhat negligible)Pros:
>Literally everything else, it's so superior that there's no competition imo>>8641999I'm an illiterate retard so I don't know what "underbaked" means, but lacking stabilization and the comparison with 1.5 and pony does track with my experience.
>>8642002whats your model of choice then, sir
and before you meme arrow, it's not perfect, but the artist replication is good, and the colors are not completely fucked which is all that is required
>>8642011https://huggingface.co/minaiosu/arca3514236951/blob/main/naiXLVpred102d_final.safetensors
Funnily enough
>>8641981 is on 102d custom, if we're talking how good shitmixes are. Still a lot of weirdness in them, but it makes sense considering it's 100% eps loras merged onto a vpred base.
>>8642003>I'm an illiterate retard so I don't know what "underbaked" meansWell, just undertrained. Prompt an artist on base noob and you're likely to get a vague red blob for whatever reason.
I think it might've been some error in the end of training too, since even a single early epoch of a lora tends to stabilize it majorly. Not to usable degrees but it does fix the worst of the weird color blob blowout.
>>8642014>Funnily enough >>8641981 is on 102d custom,are you prompting "traditional media"? it always makes 1.0 2d blow up for me
>102d
Shitty gook mix from arca where all the most terribly inept genners congregate.
>291h
God mix.
>>8639737100% yes, but mine didn't arrive until I turned forty
>>8642015Nyo!
>>8642020Cattower and Cyberfix my beloved :333
just let people use whichever model they want. you'll turn this thread in another /hdg/ if you keep up with your retarded console wars
>ask for model rec
>only receive recommendations that're in bad faith
its already /hdg/2
wtf are amazuyu tatsuki and mitsumi misato the same people
>>8642037oh shit they worked together
now i don't know if i should rebake them together or do him separately, there's some overlap
>>8642036Someone please give me good faith recommendations
>>8642041reject this idolatry
alhamdulillah
>>8642041The good faith advice is to try out all of those to see which one work best for you and your artist mixes
So there are no new models?
>>8642104be patient saar, they are making plans
>>8641614how would you go about it anon?
>>8641589definetly
Does it generally take this long to train a DoRA for 3k steps? I'm using kohya_ss.
I haven't trained anything after vpred stuff dropped so no clue on what optimizations are available now since I'm recently got a 5060Ti just for image gen.
>>8642117I don't even know if you're supposed to train Dora for that long
>>8642123>>8642122Something is definitely fucked when using my old config. Testing the early epoch lora also gives me complete noise garbage.
>>8642104why would there be
>>8642145>looking at loss graphsdented poster
>>8642104everybody packed and went home after realizing that they lack the skill, talent and means of the NovelAI team.
test
md5: 13368923b974ea9bcfc32fb05dc3a033
๐
>>8642114Same as every inpaint. Prep the area with some blobs, trying to match the color and rough shape of your intended result. Then use low denoise around 0.5, optionally add lace or skindentation to the prompt.
>>8642104rouwei v0.8 released two weeks ago
its bad but its new and trained on a cluster!
sdxl btw
>>8642234if is still sdxl how it is new?
>>8642104NetaLumina is the only thing that looks like it might turn out decent.
Kohya_ss scripts are so fucking crap, I've been changing every fucking setting now trying to get a functional V-pred lora. It should just be enabling the "v_params", and it should at least work, no?
>>8642279v-param and ztsnr only work in tandem
>>8642279Yeah, I have the --zero_terminal_snr flag set. Not sure if it's the noise offset or min SNR settings. My old settings used debiased estimation noise instead of SNR.
>>8642300>noise offset or min SNR>for vpredthe fuck are you doing nigga
>>8642259because hes training it in the big ol 2025
>>8642302You know they did use it on v-prediction models in the min SNR paper... right?
https://arxiv.org/pdf/2303.09556
any big naiggers good at vibe now that v4.5 is fully released? haven't used it since v3, not sure how much has changed. I see people talk about a negative trick and using up to 8 images.
anyone ever train a lora off of a nai artist mix
>>8642350one, yes
why do you ask?
>>8642352planning on doing it myself
probably just gonna yoink a bunch of tags from danbooru posts and feed it through with the mix
>>8642350It was a bit messy, the guy whose gens I took didn't inpaint fingers or eyes.
>>8642220h you are GOOD but I could make mine look like that as I'm quite new
>want to gen leotards/swimsuits that are the opposite of highleg
>but lowleg practically doesn't exist as a tag for this type of clothing and so doesn't work
>using prompt editing to go from buruma kind of works, but also not consistently
Another one for the list.
>>8642428for the inpainting list
>>8642433It's more of a wishlist for future models/loras and for reference about what current models can't do.
I already inpaint/controlnet/img2img so there's no list there.
novelai does it with ease
>>8642058That's like saying "just draw what you want" when someone asks for advice on how to learn how to draw
Yes technically true but it's the single most useless piece of advice ever
>>8642444You're exaggerating. It's still a decent selection, of a dozen models at most. Not too much to test on your own, and better than trying random Civit trash hoping to strike gold.
>>8642448True enough I suppose
Just that a couple of advice in here seems to go that way
"just do whatever" lmao
>>8642444>>8642456Not really, he asked for a model recommendation not how to do x or y, you can do almost the same shit on all the current models, the only real difference is how easy and how consistent the model is at pulling off certain settings and styles thus my advice of just trying out the popular options
yea bro just use SD 1.5 with a lora its basically the same
there's sd 1.5 models being released on civitai as we speak. the architecture hasn't died
just finished gta san andreas again
man what a transcendant game
Any cunny enjoyers still around here posting catboxes? I've baked a new cunny lora for Mochi (RIN) if anyone's interested.
>>8641880>Model: 0.5(NoobAI-XL-Vpred-v1.0+v29b-v2) + 0.5(naiXLVpred102d_custom)Is that the replacement to.. 291h? G-give it to me.. t-that thing.. your Dark Soulโข.
>>8642562>Mochi (RINpost a boxed example, I'm not sure I know who that is
>>8642564nta but most likely mochi_(circle rin) on danbooru
>>8642564Real artist:https://files.catbox.moe/n6z4ce.jpg
Lora example: https://files.catbox.moe/c7g8of.png
Doesn't seem too fried to me, but I don't bake enough to test Loras extensively.
>>8642565>mochi_(circle rin)oh yeah I see, nice style
>>8642572as much as I like this loli miku and her little feet, I don't think it looks like the artist's style
>>8642583I think the style is pretty weak since I for some reason was testing out baking without artist tag for the style ages ago, so the style is smeared out over the rest of the tags. Should probably rebake tomorrow with the tag again.
>>8642588would be nice, this artist has a lovely style
>>8642572even if it doesn't look that much to the artist, it looks good on its own, i'm interested
>>8642572>[masterpiece, absurdres, very awa, best quality:20]For me? 100
amazuyu tatsuki https://files.catbox.moe/jz9fvx.safetensors
hashimoto takashi https://files.catbox.moe/4yotj0.safetensors
hiiro yuki https://files.catbox.moe/um8g0y.safetensors
iwasaki kouji https://files.catbox.moe/hwyc1e.safetensors
mitsumi misato https://files.catbox.moe/iisf0t.safetensors
man, NAG (an actual working negative prompt) came out: https://github.com/ChenDarYen/ComfyUI-NAG
and reforge is dead and it won't be ported ever, that's fucking sad.
fuck using comfy.
file
md5: 49011e357f9a1a2d30429c77e009b101
๐
>>8642771its shit. barely works with SDXL
>>8642771explain how this would improve your gens without sounding like an idiot
>>8642782>>8642783did i trigger a "no neg schizo cult"?
>>8642787alright, alright, use that to remove asuka hair accessories with pure negative prompt
>>8642771I saw panchito over on reddit the other day while searching for deepseek presets. He seemed well and doing much better for himself without our chicanery.
>>8642771Comparisons with normal negs and negpip by a real user?
How do I get the pubes to match the character's hair color? It always ends up as black for some reason despite having "exact color" hair as a prompt.
>>8642876You're saying tags like "orange pubic hair" don't work? Have you checked if it's a artist/lora/shitmix issue?
>>8642878>Have you checked if it's a artist/lora/shitmix issue?Ah right I haven't checked if one of the loras was fucking things up. Ty anon
local will never recover..
>>8642884That's okay, I use NAI, now post the metadata.
>80 lora datasets
Man, I feel like I'm in 1.5 times again.
>>8642884>>8642896NAIsissies.. our metadata?
>>8642911sorry fellas its my secret powermix + vibe
>>8642911beta you should never let your trade secrets known to others when you could be making bank on patreon...
>forge classic makes gens slightly more local lampier and shinier on the same exact setup
I wish I understood the mechanism behind that kek
But I also ought to finally fuck about and try debloating reforge for fun myself
>>8642986>debloating reforgeI just unticked half the extensions that came in and it's already much more usable
>>8642991Well it's not bad when you get it set up but it still does a bunch of useless shit and installs like 310 garbage modules that nobody ever uses. Plus classic has a lot of new launch flags and shit, better loading times, no leaks, etc.
But I ain't gonna switch to something that gets me worse gens for slight convenience
Any of you know what art style combo this guy using? Cyber-Wifu https://cyber-wifu.itch.io/
>>8643018now that's a level of grifting rarely seen. selling gens of your wildcards for 10$ a piece is something else
>>8643023Are you telling me this masterpiece isn't worth the 5$ price tag?
>>8643018Just buy the raw image and check the metadata, lil blud
>>8643027I checked kemono. Found this, but can't figure out if metadata still inside it or not https://kemono.su/gumroad/user/4611555991622
>>8642501https://files.catbox.moe/a6vheq.png
https://files.catbox.moe/lzevhe.png
lora from anon
>>8641726
yup, another comfyui classic, I especially love the random blue artifacts on the strawberries, really shows how much of a powerhouse it is for expressing your mind's eye on your screen on a level that no other tool can compare.
657
md5: bf812abd43395206ef67f8c54c90a6f3
๐
this is nice
although i had to comment out print commands out of like 20 scripts
>>8643092my favorite part is the elbow nipple
>>8643076What did evaanon mean by this?
>>8643095And looking at it again, the random blue spots are pretty much everywhere, must be the new AddLatentMoldINversionGPU node.
>>8643076Was this supposed to be pixel art? Or did you just fuck up the latent upscale so badly it ended up looking like that, random inverted color spots and all? I honestly can't tell.
>>8643100LMAO it really does look like mold on cheese https://files.catbox.moe/kmuv4m.jpg
>>8643100>>8643103judging by the overly saggy tits, it was meant to be a shitpost and not something to be analyzed in detail
i feel like i've also seen it before but eh
>>8643103it does have some oekaki and pixel art but i dont know where the blue artifacts are coming from. Just messing around with different stuff
file
md5: 826daa6c4c79e57efb0bdcf047727474
๐
it seems like its some scheduled loras not playing well with controlnet tile upscale, the base image is fine
>>8643132It doesn't make sense to schedule a style halfway through when you're starting from an image that already has the style.
>>8643139i only use it to get a certain anatomy i like so i cut it off at 20/30% so its kinda like img2img since at that point its still a blurry silhouette.
>had a fully working scraper first time
>deleted it
>now gpt can't replicate it for its life
fucking useless
I have trained a model on illustrations I've commissioned from an artist, it did pretty well.
>>8643176... you trained a lora or a full model?
>>8643157>got it back using photorecterrifyingly effective tool
anyway i wanted to take another look at artists but smartened up and made it scrape the top pic from the last 35 uploaded ones of an artist, and lowered the threshold to the same
maybe it'll make more sense
>>8643059Ah okay. So the lora mixes well with other artists too.
>102dfinaltraitor
>>8642302>>8642311I really don't want to let this slide. Wtf is the answer here? Thread has always said turn off your noise offset and enable those two settings to train vpred, then, as usual, someone appears with official source saying he's wrong. Who is right here? I don't give a shit about theoretic bullshit, what actually works in practice?
>>8642350Several dozen.
>>8642354Set up ranbooru with nai extension on reforge and go to town.
>>8642854Make your own, sweaty. Just workshop with deepseek in a character creator card what things you like and what things you don't want it to do and you should be good.
>>8642876Orange pubic hair (though this might make the guy's hair orange). Mispatched pubic hair in negpip would be far more specific to each character and versatile.
>>8643130For me this was from using a eps/illu lora on vpred.
>>8643330Yea, litter seems to work
>>8643330Seems so, image load but there is no way of accessing the upload interface.
>>8642787no, i genuinely tested it, and it did more harm than good.
Why is holding a fucking clipboard so tricky to do right
>291h
>h
>102d
>d
It's very simple. One mix for each board.
>>8643354no gen for either tho
>>8643358Gens are in /hdg/ though. Don't want to interrupt important lora training debates with useless distractions.
>>8642592>>8642599Tried rebaking with the artist tag, but I don't know. It looks like mochi, but I feel like it's burned in lightning since dark barely works with the lora. Starting to understand the lora autism anons here, since I've rebaked 4-5 times already. Still undecided on if I should train unet only.
https://litter.catbox.moe/8eae7p1zhsp9rkx3.png
>>8643408unet only was always worse for me
>nyo gen
mmm how about, nyo?
:3
>185 artists scraped
Yes I think I need to only choose a few of the best
>>8643408hard to say, try to make a normal neutral face looking at the viewer, so it's easy to compare
>>8643354My nigga.
>>8643358Not my nigga.
>>8643309he's a retard. disable noise offset and use either min snr with gamma=1 OR debiased estimation + scale v pred loss like noise pred OR edm2 OR neither of these at all.
>>8644198So you're saying he misquoted the source?
>>8644198This is correct, noise offset is the only one that's nonsense if you're using zero-terminal SNR. The other options will work on V prediction models just fine.
>>8644201min snr paper never said anything about noise offset. min snr weighting scheme is valid in either eps or vpred, however you need to apply a slightly different math depending on the prediction type (which is already handled in most trainers)
I have done this pose dozens of times and I am never sure if I did it right
>>8644357I think I had that problem when I prompted for "split" instead of "sitting split". There's also "side sitting split" on danbooru which is more specifically this type of split, but I haven't tried it to know how well it works vs just "sitting split".
>>8643309noob team used v-pred and z-snr, so use the same for your loras
disable pyramidal noise or other noise offsets because they conflict with z-snr
I wish there was a tag for the floppy type of dog ears. Most of the time it just generates normal dog ears.
>>8644393Wait wtf, floppy ears actually is a tag, and it even has 10k posts. I thought I tried this wording a long time ago but maybe I'll try it again now.
>>8644396Hmm ok. So it does work. But it looks like shit with the artist mix I use, maybe that's why I memory holed it.
https://files.catbox.moe/rxm0ri.png
What Lora trainers is everyone using? I'm starting to get a feeling koyha_ss still has several annoying bugs with v-pred model, since I can only get the standard lora to train properly. Every time I try to train a Lycoris/* type Lora (LoHa/Con) it ends up completely skipping unet entirely and spitting out a 28MB file that only trained the text encoder.
>>8644404why is it so noisy?
>>8644467not sure what exactly you want to see there
https://files.catbox.moe/c0e2xr.png
i'm baking some avant garde shit
i wonder if it'll translate to porn at all
>>8644475just bake a rio lora?
>>8644476i don't know what that means
Formerly bros..
https://files.catbox.moe/v2w6uv.png
>>8644462Isn't it all GUIs for the same exact kohya shit anyway
>>8644482machina fork is based on kohya but the dev made quite a few changes to the backend code
>the more images an artist has the more boring and generic the style
tsk it really is quality over quantity
file
md5: f8aab28a04a155face48572928d74cb7
๐
almost got banned from /vr/ lol...
file
md5: 632d5b53e2acedbe2125922a55ba2887
๐
>>8642549Those eyes... Ryouta Magaki lora? Or is it Amazuyu Tatsuki?
Tell me right now, busta
>>8644566that was the amazuyu tatsuki
>>8644550what? a fucking /hhg/gger on my hentai thread? post your latest bought chinkheld
>>8644358 At least you finally got the clipboard correctly.
>>8644480What do I have to prompt to get the penis to nearly vanish between the breasts like this?
>>8644588I also did on other gens as well but I liked way more how things looked on the deadnoodles mix
>https://files.catbox.moe/1q3jbk.png
>train lora without artist tag
>inconsistent, but good when it activates
>train lora with artist tag
>almost fried in style with anatomical mess
Maybe it's time to put down the lora training.
>>8644591Usually not adding any tag for the penis plus sufficiently large breasts does it for me.
>>8644594https://files.catbox.moe/l6yfcg.png
>>8643408Well, at least version 13 of the lora makes nice looking images. Not that strong mochi style though.
https://files.catbox.moe/cniyiw.png
>>8644464It's bow_\(bhp\)
>>8644495For me? Searching for the latest gacha slut, order by artist, pick the top 3 with the most results. It's genning time.
>>8644550Based troublemaker.
>>8644591don't tag penis
https://files.catbox.moe/6w7ff7.png
>bake character lora
>goes almost perfectly
>bake style lora
>model learns some crude features and then starts frying
it MIGHT be over
>>8644759config issue. never listen to dataset fags. little bit of style grouping and eva are enough
>>8644762>style groupinghow?
>>8644597>Usually not adding any tag for the penisYeah but this sometimes makes there be no penis at all unlike what you have.
why do anons use esoteric terms to explain otherwise simple concepts?
>>8644767just kick any obvious outliers from your dataset. Not really necessary for consistent artists
>>8644774Because people love the idea of being smart, and since they are stupid and have a hard time understanding what smart people say, they think that if they use esoteric jargon people will be impressed and recognize them as the smartest boy in the classroom.
>>8644774People synthesize information in a way that makes sense to them, so just evoking a few sentences will call all their previous ideas and premises. This is great for rationalizing information but sucks for communication
>>8644784You're doing what you're accusing others of doing anon, it's not that complex lol
for me it's random idioms that only make sense for me in my head
>>8644775I'm mixing doujinshi pages with normal illustrations, is this what is causing my hands to deform? Problem is that I'll only have 20-30 (from 140 rn) images if I prune everything that's not the most consistent style in my dataset.
>>8644824>is this what is causing my hands to deform?could be. can't really say without seeing the dataset and config. too high lr for the te is also a prime contender
How do I merge checkpoints?
There's this one checkpoint that I absolutely love, but it's somewhat ruined by a slop snakeoil style lora the author added in
He left a recipe in the checkpoint page
>The main checkpoint he used as a base
>Epsilon pred at .3
>His own lora at .5
>A night lora at .5 to bring the saturation down
How do I make use of this? I've never merged checkpoints before
>>8644835Privaty is for anal only.
>>8644835privaty my beloved
>>8644827I appreciate the help anon. Here's the latest config I've tried: https://files.catbox.moe/pr429q.json
I'll probably upload the dataset if the next few tries don't work out.
>>8644844did you upload an empty file?
>>8644847No? Guess I'll try litterbox instead: https://litter.catbox.moe/poacfgvnswgosdgu.json
>>8644838Just cause she has a fat ass doesn't mean she's anal only.
>>8644841
>>8644855It's her fat ass and tsundere bitchy personality and the fact that she is a glowie.
>>8644608looks nice, can you upload this version of the lora?
>>8644859Sure, here you go. The style is working, but hands and the males are coming out a bit deformed every other gen for me. The males makes sense since all of them in mochi art are basically monsterous ugly bastards, so that's a expected, but I don't really know why the hands are ruined.
https://litter.catbox.moe/rfbuq3l6i4bvc11w.safetensors
>>8644885thank you. is
>>8644824 you? I'm still too inexperienced with lora baking to say anything for sure but I think since v-pred 1.0 does have some idea of what the artist looks like you could choose only the best images and train the artist tag (instead of morin00) to reinforce it instead and hopefully that would clear up any issues
>>8644858>and the fact that she is a glowie.Haha, definitely a ridiculous part of her personality. Got back into Nike and I'm still not sure why she stans them like she does.
>>8644891>>8644852No clue on how I migrate from json to toml in koyha, but it should contain the same stuff right?
>>8644893I tried training on the real tag but got poor results, since the tag is a bit of a mess due to mochi* being shared by several artists. Making up a new tag seems to work better, ideally I would have a new embedding but I can't be bothered to do the whole textual inversion + lora at the moment.
22 done, 6 datasets to go
>>8644852not really familiar with args on kohya, but 0.1 module dropout plus dora seems like an odd combo
>>8644911I didn't even know I was training with DoRA enabled... I thought you had to select some other type instead of "LoRA_type": "Standard" for that.
>>8644915dora is just a way to decompose the weights
>>8644916Yeah, but I've never seen the option unless I select a LoCon/LoHa/etc in the gui, so I had no clue it could be enabled for the standard.
>>8644905Box? I've been trying to figure out this position.
>>8641675Yeah I hate ai gloss
My only issues with 291h so far is that it lacks details a little bit but that cfg scale boost lora solves this issue for me
>>8641137>>8641675>>8644954we use naiXLVpred102d_final here
go back
>>8644928I had to doodle the buttplug-tail and inpaint, but the base pose is
>facing away,implied kiss, hug,grabbing another's ass
>i hate gloss
>so i use the shitmix with literal 3dpd and furry shit
based desu
>>8644962>finalYou are greatly encouraged to make a hasty retreat.
>>8644975>implied kissHmm thanks a lot.
if you suffer from gloss on any of those it's entirely a skill issue
>>8644995Yeah nigga I have skill issues
Now either solve them for me or go do something productive
>>8644997I'm doing gens :)
I feel like this is gonna start console wars in the other thread so let me ask here and bear with me, I'm asking sincerely.
Does Pony have any strengths over illustrious/noob based models? I can only think of style training being easier, but it seems many people still use Pony merges.
Also what exactly are vpred models better at? As I understand it it's better color range and lighting prompts?
>>8644975What settings do you use for inpainting? It looks well integrated there while I feel my inpaints could be better.
>>8645016>Does Pony have any strengths over illustrious/noob based models?not really. guess it has more loras
>>8645016>Does Pony have any strengths over illustrious/noob based models?Much to my chagrin it is much better at 3d, only thing I can think of
>>8645016It can do better fully realistic photos of Pony figurines
unironically
>>8645016I guess if you are more into Western stuff, but Pony is literally burnt to shit and I really don't know why you would use it.
>>8645016>Also what exactly are vpred models better at? As I understand it it's better color range and lighting prompts?In theory but it also makes all anime look dogshit
Eps isn't better because nobody shitmixed it properly yet though
>>8645017I said "inpainted" but I didn't actually inpaint, I did some basic redrawing on the base gen and then latent upscaled by 1,5 with 0.8 denoise and Tile CN with 0.35 strength and 0.75 end step, and then did a second latent upscale by 1,335 at 0.75 denoise, same CN settings.
>>8645021If NAI can do figurines well, why can't noob?
>>8645016Take it from me since I'm a huge casual and a complete retard
Pony is better in realism, in overall. This means it's better at genning all manner of 3D than Illustrious/Noob.
Everything else is SHIT.
Prompt adherence, the need to use score tags, the shitty inconsistency, the lackluster database, etc. The only reason it has so many loras to begin with is because it was so bad, so many characters and styles were missing, and you can't really use artist styles since they're encrypted by some esoteric, autistic reason.
>>8645013Already did
>Does Pony have any strengths over illustrious/noob based models?Not really if what do you want do to are anime related gens
might as well post this here
asura lora, use the artist's tag as the trigger https://litter.catbox.moe/jsfigv.safetensors
trained on v-pred 1.0 with lora anon's config
>>8645053>some esoteric, autistic reasonThe reason is that Western artists are extremely anal about copyright and since it's mainly an /aco/ model Asstrakike ashed the artist tags. It's annoying but I understand why he did it.
>>8645060I remember it being more than copyright but AI as a whole. Astralite found himself in that weird space where he was trying to please westerners who hated AI inherently and would never be happy or proud of him regardless. Hashed artists was just the final nail in the coffin and guaranteed he'd be dethroned as quickly as possible.
>>8645055>If you're getting glossy stuff it's your fault>Posts gen>It's flat color animeC'mon man. The glossy issue surfaces mostly on styles that tend to 2.5d or oil painting.
That said, while I've never used naiXLVpred102d_final, 291h is remarkably hard to get unnecessarily, unprompted glossy gens like most models on civitai do lol
>>8645016The main advantage of noob v-pred is that it's the last version so it has the strongest danbooru knowledge. Color range and lighting benefits go away as soon as you merge it with something else or apply a stronger lora, which most people do anyway.
Pony is still somewhat better at photorealism when used in merges. And it's the only model so far to be trained on R34, which means /aco/ characters. Think Dexter's lab, Drama Island, WoW, MtG, etc. And /aco/ styles, the few ones that aren't hashed. Many of get pruned off danbooru entirely, like Magnetus or Deuce.
>>8645067Not my fault you have missed all my previous gens with other styles using those models
i need your strongest noob 1.0 vpred toml
>>8645067You should be using custom, not final. Now please, retreat.
>>8644983
>>8645040Yeah I need to get back into trying latent upscaling. Thanks for the settings, makes sense it looks clean.
>>8645057why bother stripping the metadata
>>8645067It's not like glossy is something AI brought along https://exhentai.org/g/5582/10ab141549/
But knowing AI shitmerges struggle with it makes it feel like the person failed to wrangle their model, not like he made the intentional choice to include it.
>>8645078>You should be using custom, not final.why? Final is better and is artist friendly
>>8645079Latent upscaling has the disadvantage of adding lots of noise depending on the style (at least on noob vpred), so I pass my gens through https://openmodeldb.info/models/1x-ArtClarity between steps and sometimes after the second step. It's not perfect, but it helps.
>>8645082retarded paranoia, my apologies
>>8645085>complaining about 2.5d shine>decides to use the shiny model anywayIdk what to tell you.
>>8645073Already posted itt. And not like you need anything extra for v-pred, it's still the same base arch. Use regular SDXL settings except
>>8644392
>>8645087>retarded paranoiaWill a jeet steal your super special settings??
>>8645072I don't know how to distinguish between autistic schizos online
That is pretty good though, although it's still very anime heavy
>>8645078Not that anon, and I haven't used final or custom, just a bunch of shitmixes from civitai and 291h
>>8645083Agreed, I think it's mostly that people tend to like the glossy thing and it's been so overdone that it became a slop trademark
>>8645095For me I use custom as my main, then if it smooths out and artstyle or lora too much for my liking I try 291h. In most cases 102d custom is better.
what are some good artist tags to make your gen more /h/
like more hentai expressions, dynamic posing and just sexier stuff
hard to explain
>>8644474where can i get that model?
>>8645092anon posted them here
>>8633915
00086
md5: 8bdf6dc06c4b90845c97630282208014
๐
>>8645095>although it's still very anime heavyYou see, I like anime and that's what I like to gen, very weird given the board and thread topic I know
>>8645104just... use a lora
unironically though noob artists are FUCKED!
>>8645104those can be solve with mere tags, check out the danbooru list for those
>>8645098Better in what sense anon
We all know that word is subjective when it comes to this so I need a more precise description for me to trust you
I have a preference for more realistic, painterly, dense artstyles like this
>>8641961>>8645123>I like anime and I like genning anime>This gives me the authority to speak on less/non anime genning
>>8645132NAIchads.. have mercy
>>8645135>I have a preference for more realistic, painterly, dense artstyles like this1.0 2d custom sucks at those unfortunately
>>8645135use final, custom is lame
>>8645132yeah I think I'm starting to get to that point, artist mixing can only take you so far, things just kind of feel rigid
any good loras you'd recommend?
>>8645133obviously I tag for the shit I want to see, but sometimes you're just stuck with artists or an artist mix that just can't produce what you're after, or at the very least not do it very well
>>8645135>We all know that word is subjectiveI mean you've answered your question already. What I will say is what I've said already. 291h gets the details that 102d smooths out to improve anatomy. Which reminds me, I need to make a lora. If it comes out good I'll post a pic.
>>8645086Yeah I used it in the past but ended up going back to a non latent upscaling with animesharp v2 soft at 1.25x with the noob anime-lineart or manga-line CN at .45 denoise .65 strength and 0.55 end step. I also do my base gens at a higher size than the standard sdxl resolutions. Gonna try your strategy though, looks good.
>>8644572i dont even have a chinkheld... only my [spoiler]phone[/spoiler]
I really like "some" artists on base noob v-pred
enough to use it once in a while
>>8645143Too bad
291h gives me decent results with these (like that gen, which I just threw together to test the model)
Sloppier models (like NTRmix) give me greater quality overall, but then I'd be stuck with a model that has bad prompt adherence and knowledge and consistently gives me glossy AI shit lol
>>8645146What makes you say so anon?
>>8645150It's subjective in the general sense, not in the inner particularities of the topic
If you say A model is better than B that doesn't tell me much, but when you say "it's better than B at doing X", then it's pretty objective desu
>291h gets the details that 102d smooths out to improve anatomyWish I had the hardware to use refiner
>>8645147mine ha ha
but yeah its not even rigidity or anything but vpred just fucked over anime
>>8645135Alright, I consneed, that style is not something I have ever tried before
>>8645147>but sometimes you're just stuck with artists or an artist mix that just can't produce what you're afterWhen something like that happen to me I just switch to another mix, sometimes you just need to accept that perhaps your favorite mix is not going to do well with everything you want to do
>>8645155shamefur dispray
>>8645165>Too badyes... strange coincidence but yuming li is one of the artists I plan to make a lora of in the near future
shuffle captions or not for style lora??
>>8645183shuffle captions for every lora, and just keep the first tag for characters
>>8645183shuffle captions for every lora, and don't keep the first tag for characters
>>8645183shuffle+tag dropout at 0.05-0.1
Anyone have the catbox/style for this? I haven't been around in a good bit before this week but honestly, this is hands down the sexiest/hottest style I've seen.
Hell, I still have the thread link from back then if that help: https://archived.moe/h/thread/8258147/#q8259629
>>8645186>>8645188>>8645189I'm reviewing configs posted here and some people aren't shuffling captions. His lora was good so idk. More stuff to test. Not sure why there's so much dogma here.
>>8645181Love his (or hers?) stuff
Also artists like hungry clicker, kornart, jonpei, etc
>>8645193the big redpill about loras is that as long as you don't majorly fuck up with 1lr or anything it'll probably bake okayish
why do you think there has never ever been a consensus on stuff other than whatever the first guides said and just stuck
>>8645199>why do you think there has never ever been a consensus on stuffBecause people will call you retarded for every little thing even if you post proof.
>>8645205eh it used to be normal in them hay days and it was still the same
there's no one good way to do this and i say it as the testing autist who's been testing since then
>>8645191hair looks like quasarcake
What's are good hi-res fix settings for 291h
>>8645227>minor spelling mistakemfw
>>8645227you don't need different upscaling settings for every shitmerge
>>86452271.5x 0.4 denoise dpm++2m
Thank you, whichever anon suggested Euler SMEA Dy + beta for 291h
>>8645227cn tile upscale
>>8645298No prob. What does it do that other samplers does not? Curious about your case usage for it as I have my own.
>>8645354What's that first one? Sorry, I'm very much a novice
>>8645354What I've noticed at least with euler and dpm++ 2m is that the genning isn't uniform
It spends most of the time as a dark blob and only on the last 40% of the gen it actually gens something
This setting makes the genning uniform again, so I can finally do prompt editing normally
>>8645381>somehow grainy and blurry and oversharpened at the same timehats off, and the disgusting hyper anatomy is just the cherry on top
>>8645396thank you, it will happen again
i give it like 1.4 days again
>>8645214It's got nalia too. if anyone has this anon's gens or he himself is here, please pass a brother a bone.
>Does Pony have any strengths over illustrious/noob based models?
>I can only think of style training being easier
>>8645411deliciously suckable thighs
What front end is this?
which ghost is lil blud fighting this time? :skull:
which ghost is lil blud fighting this time? :skull:
>>8645428Front end as what ui? comfyui
https://files.catbox.moe/izt2x6.png
>>8645447Artist is airchin_doggo. Just prompt the composition.
>>8645447https://files.catbox.moe/4nj4dp.png
sorry
>>8645462>NoobAI-XL-Vpred-v1.0+v29b-v2.safetensorsshould have guessed from the background. Thank you.
>>8645466>should have guessed from the backgroundBased.
is the metachs nai extension not updated or have some issues? I'm trying to inpaint with 4.5 but it's not infilling, just giving some basic low denoise img2img, v3 works on the webui, and 4.5 works fine on the nai website.
i wish there was a good way to merge rouwei and noob vpred 1.0 together
i did try it, got the character knowledge and styles got unfried with rouwei in, but it became extra gray and anatomy is subpar. if only there was a working way to extract a lora from rouwei and just use that with noob. but doing that only gens pure noise.
>>8645562>>8645562Can't say anything about the anatomy, but these saturation and contrast loras are top tier:
https://civitai.com/models/1093089
They have no effect in the style, composition, character and all, it's almost like using a filter.
>>8645565oh, thanks. i guess that's one way to bandaid fix it
>>8645566This might help with the anatomy
It basically mimics the effect of having higher CFG without fucking up your gen, so at lower weights it helps with coherence
https://civitai.com/models/99619
While we're sharing all these useful loras, don't forget the one that turns your gens into henti
https://civitai.com/models/341353/expressiveh-hentai-lora-style
Does anyone have a lora for inpainting sneakers? No fancy HD realistic collector 1965 edition whatever the fuck. Just simple and clean generic sneakers.
>>8645016lipstick mark are really good on pony and sucks on ills
>>8645627Oh I remember you. Hello anon. To be fair, "lipstick rings" doesn't even work on pony so I'd say they're even.
>>8645628first time I brought up that point so not sure about your schizo theory, rings doesnt work but smear and overall smudge is way better than on illus, not sure why
>>8645641There's one anon in particular who loves to prompt lipstick mark. I guess it wasn't you.
>not sure whywesterners do this better
bruh, has civitai started to throttle downloads or is it just slow as shit right now
I'm downloading at like 3 mbs when I should have at least 10
>>8645771ok now its picking up
shit site
anyway what other popular stuff besides
291
102d
base
eih/sih
is there, i want to look at some shitmixes today
You've tried the rest, now try the best. Redeem CatTower, NTRmix, and WAI. Enjoy henti.
>>8645867291 already has all of those sweaty
>1.0 eps base got me a better result than 1.0 vpred base with two vpred loras
jej?
>>8645872i mean eps base with two vpred loras and vpred base with two vpred loras to be clear
What caused it this time?
what cyan i say
https://files.catbox.moe/jp3rim.png
i was surprised that it didn't explode, thought something baked with vpred flags on vpred1.0 wouldn't even work on eps at all but at least here it's better
tho ofc it needs more than one xyz
my highly scientific conclusion
if you're using style loras the shitmixes have almost zero difference between them
test
md5: f9f6a0a4ffb54c992623f3a5d81ebd6d
๐
>>8645864All you can expect from merges is style bias and different levels of knowledge dilution. Base, one or two of the others and WAI cover just about every use case other than really niche stuff. I can give you some of those niche ones:
>hsUltrahdCG_IllEpic - heavy 3D bias, helps with Koikatsu/DoA/Blender looking stuff (with lora/artist tag)>pornmasterPro_noob - uncanny valley half-realism, for setting up img2img into actual realism models (unlike those it maintains most booru knowledge)>PVCStyleModelMovable_nbxl - plastic figures that work with anime artists and loras
>>8645909>PVCStyleModelMovable_nbxlThat's funny
>forgot to remove the meme quality tags from pvc model when genning regularly
>image like 10 shades greyer than without them
hm
>gazillion datasets
>only baked a portion
>only like a couple of those
Oh well sometimes you get stinkers
butakun https://files.catbox.moe/omk4vz.safetensors
eencya https://files.catbox.moe/ytxscz.safetensors
eitopondo https://files.catbox.moe/k8fonr.safetensors
fujirin https://files.catbox.moe/4vh8mn.safetensors
godiva ghoul https://files.catbox.moe/ehaisp.safetensors
henyaan https://files.catbox.moe/vzzct5.safetensors
kirihara you https://files.catbox.moe/p0kyy8.safetensors
nisino2222 https://files.catbox.moe/xo1bex.safetensors
nori tamago https://files.catbox.moe/8yezvb.safetensors
shiozaki yuji https://files.catbox.moe/0ub0hp.safetensors
watarabe keiichi https://files.catbox.moe/wvtlwa.safetensors
eencya came out pretty neat at least
Why is there /hgg/ and /hdg/ if they are literally the same?
The OPs don't even specify what the difference is supposed to be, so naturally you just get everyone posting in both.
>he says as 10 loras were just posted here after a model discussion while hdg is seething about some personalityfag for the last two hours
>>8645961Cool, non of these are painterly though.
>>8645930I never even read the description, it just works with regular prompts
>>8645983there's no difference other than trolls and spammers are too lazy to squat in both so they stay in the more active one
I wish we would merge them back together, it's a pain having to check two places
>>8645961>eencyaHuh. Is it good at generating anything other than Ame?
>>8645961>another big lora dump>no xerbattpain peko
>>8646005It ain't bad
https://files.catbox.moe/zyolaz.png
also fucking hell japs are boring
ive gone over 25k danbooru artists again and i got a whole 3 i want to use
oh well, most of artist i like are acoshit that don't look that good outside of their images imho
>>8646023give me a dataset and i can bake it pretty quickly lol
>>8646031oh sure, give me some time
>>8646031>japs are boringBake western painters. Japs are subhuman anyway.
>>8646039I've baked Rubens and Van Gogh before lol
They worked SURPRISINGLY well
>>8646031did you do Nanameda Kei already?
>>8646041Well just keep going. At least these styles are unique and worth something, jap smut is cheap and mostly looks the same.
>>8646041Yo really?
Could you do Bouguereau?
>>8646046DESU I am more into classical than paintings, I wouldn't know what to bake.
>>8646041There's two Rubens loras in the rentry
Which one is yours?
>>8646048Why not
Rubens worked well enough to show the foundation under chipping after upscaling even though it was baked at 1mp
>>8646051Probably both from rebakes
But it was with my old shittier setups, I could rebake it on the newer one
>>8646056Oh, cool then
I'll wait for you to do it
>>8646050>>8646052Bouguereau is pure kino
One of my favorite artists, although admittedly I'm more of a golden age of american illustration fag (Leyendecker, Norman Rockwell, NC Wyeth, Howard Pyle, etc)
>>8646060Do you know any good HQ repos for art? I was hand scraping wikimedia before for the two. Archive.org?
>>8646066Wikiart is good for finding the specific paintings, you can reverse image search to find the higher quality versions
>>8646066Try this
https://www.reddit.com/r/ArtHistory/comments/tjhdo3/comment/i1kttm5/
>>8646066https://artrenewal.org/artists/william-adolphe-bouguereau/7
>>8646071>>8646078This will probably take a long time doing by hand.
>>8646066Bouguereau in particular is probably still easiest from wikimedia commons? It has this neat page with all the artwork, and most of it is in high quality too (I guess it's a famous artist): https://commons.wikimedia.org/wiki/Paintings_by_William-Adolphe_Bouguereau
But you'd need to write a script that goes to the link in each thumbnail, finds "Original file" link there and downloads from it. If you don't want to bother with hand clicking all that shit. I can write the script if you need help or ask AI or maybe you can do it yourself, idk your technical level.
>>8646084I mostly ask because I noticed Wikimedia didn't have all of the stuff from Rubens on their site, or in the highest quality. I had to go into his actual article and look for stuff.
But I'll fuck about with GPT for that script, I'll let you know if you need to bother.
The Van Gogh lora is great
While I am also tired of seeing shitty jap slop, it's not really just a race issue. Humans are boring. Only few have the ability to create truly great and unique art, and Japan is just a tiny country compared to the entire world, so the statistics match up. People are just weebs so we end up spending more of our time looking at their shitty art than our shitty art.
>>8646100I don't think anyone made it a race issue anon
I'm super racist but this is very obviously a cultural question, it's not like western art hasn't declined to the point where the good artists of this century are all self-taught
>>8646100Well I would say westoids make more interesting art, their anatomy is just too western to enjoy
Japs are the opposite, nice abstracted anatomy but like 95% of their art is 1gacha standing
I thought it was a model problem when I was prompt sring danbooru artists natively but well now I know different
>>8646103Bit of an unfair generalization, seeing as "western" art is literally everything outside of china and japan. From actual painters, cartoons, 3D all the way to people pretending they make anime. Anatomy similarly goes from photorealistic to powerpuff girls.
>>8646104Obviously I am talking about hentai here not the Vienna Museum
I got the script for wikimedia, now I'm trying to get a script to push artsandculture links into deezomify-rs
>>8646102>I don't think anyone made it a race issueBro
"japs are boring"
"Japs are subhuman"
>>8646103>westoids make more interesting artAnd what is the western population and number of different cultures compared to Japan?
>Japs are the opposite, nice abstracted anatomy but like 95% of their art is 1gacha standingMaybe because you are looking at a single country's coomer artists and models trained mostly on that coomer art. Of course it's going to be biased. It's just a single small nation. Pretty funny how the global audience gives them so much attention.
I don't get above guys point desu
>>8646110Just gotta say that your loras are of remarkable quality
Hopefully you're around when I get to learn how to make loras lol
>>8646122Which? If you mean the classical ones then like I said, older shittier setup, maybe they'll be even better when I rebake them.
>>8646127Yeah, the classical ones
Hope that I get to make some style loras like that too
Uegh fucking hell, dezoomify works really well for scraping but from what I see it only takes single image urls from the google site but ofc they don't really make them easily available for scraping. Hm.
>>8646131>can get a scrape using a headless browser script>but only for 33 images because it only loads 33 per scroll>gpt can't seem to get how to do it otherwiseNghm.
>>8646151Glug-glug-glug.
Phew, that was tasty. Now which jap artist was that?
>>8646155My boy Imai Takahiro.
>>8646151>Nig handwas almost good
>>8646158That's what makes it perfect :3
>>8646157Girls be like:
>me? ha yeah man I'm good>just a little tired
>>8646084Well, if you want to fuck about and try to force artsandculture to give you the links to all the artist assets for use with dezoomify, be my guest, I cannot get it to work. Consistently cockblocked at 33.
The commons script was like two gpt instructions, at least. https://files.catbox.moe/2hcx1z.py
file
md5: 0f324a44d06ece1654b829c58bd4ad63
๐
I also just realized that for some reason https://commons.wikimedia.org/wiki/William-Adolphe_Bouguereau has more paintings than https://commons.wikimedia.org/wiki/Paintings_by_William-Adolphe_Bouguereau
>>8644844>>8644827I'm back and I'm still frustrated with the Lora after 17 rebakes. Litterbox wouldn't accept the zip, so here's a mega to it (beware of cunny and ugly bastard): https://mega.nz/file/5EwGVIjC#Az8Mz086TS021t9RoL6D1gYawN-HkpLj1qLDSuFw9eQ
password: mochi
I know tons of outliers in my dataset, but I'm scared it will fry easier if I trim too much. Maybe I've just done something really dumb with the tags.
>>8646158he's clearly okinawan bro, relax
I have been refreshing board . 4chan . org for months, waiting for 4chan to be back. It turns out that URL is just not valid anymore. Fuck im glad to be back.
Is Illu still the top tier model?
>>8646169Yeah the commons is easy to scrape from obviously. I'm not too sure in my ability to scrape from google though. Since you already figured out how to get list of images into dezoomify, I assume you only need said list? I.e. all links to individual image pages.
>>8646188Yeah, but all of the GPT scripts that worked for me only scraped the first 33 results for any artist, I assume because it's a JS mess without an API so it only sees the first page.
>>8646181Asked it to give it multithreading and to work with the artist pages.
(Still unsure why the "paintings by" would have 1/5 of the images on the main page but oh well
https://files.catbox.moe/gwc40d.py
>>8646190So if you are loading from this page for instance: https://artsandculture.google.com/search/asset?q=Bouguereau
You need to scroll down first so that every item loaded in your browser, then you can verify that all the images are there in the page code: https://litter.catbox.moe/7hhsq7jyezfsxsi9.png
Now you can save this code (maybe even only the element that contains the images) and ask gpt to write a script that extracts links from it.
As far as doing it by just giving it a link to artsandculture artist page - probably not possible without Selenium or some shit like that (which automates user input).
>>8646196If you can be bothered, do it. I got more than enough from commons and trying to wrangle the jew got me an aneurysm. Bouta bake that bouguereau
>>8646198Bouguereau should be entirely fine from commons. I thought you were trying other artists.
>>8646199Well I was looking at them too just because they have seem to have the highest quality images. But yeah, fuck them and fuck their site.
>>8646201>just because they have seem to havesee i told you about that aneurysm
>>8646201Can't anything be done with art renewal center? They have pretty high quality stuff if you have an account, plus they have a large gallery for every artist and a neat section called albums where they categorize the artworks into their respective themes (might be useful if you're trying to skew the style a certain way)
>>8646158Of all the things wrong with that slop, you focus on the skin tone?
>>8646204Maaaybe but commons seems to have enough for now.
>>8646205Not about focus, he said it was good otherwise and that's the only issue
seeing all those paintings makes me want to grab a cheap flight to vienna to go to the imperial treasury again
>>8646208Cucks tend to be illiterate
oh yeah forgot van gogh was the primordial weeb https://files.catbox.moe/kjw64q.jpg
>>8646031>>8646036Alright, there you go, untagged but if you want me to tag it just say so
>https://mega.nz/folder/GnxkyZBA#ib8DCfWFnjkXL5eEnXZLog
>>8646222>262hm maybe tomorrow then because of the classical i'm baking now
but i will tag it dw
>>8646222soul but I can tell how awful this is gonna look in AI
>>8646228yeah sure take your time and many thanks for that
i am gooning to this today https://files.catbox.moe/t9nuot.jpg
>>8646237Unfathomably based
Man of elevated character
>>8646237Is this where that 20loras guy gets his inspiration from?
>20k by 20k pixel images
That's gonna be nice for the 192ch vae
holy HELL https://files.catbox.moe/tpcoc2.jpg
>>8646259wtf don't dox e like that
>>8646222Can you get that on catbox? MEGA just refuses to let me download this for whatever reason.
>>8646281A little too big for catbox so I used litterbox instead
>https://litter.catbox.moe/33x760t5h0a7nlvt.zip
>>8646284>also stops at 9mbnow i don't know if it's the previous box issues or my internet is fucking up for no reason lmao
i'll reset my router
>>8646288oh yeah it was my internet never mind
>accidentally baked the van gogh and buybububu with fucked up tagging
oh for crying out loud
>>8646237based
so much better than hentaislop
>actual literal painterly loras are being baked here
uh oh
>>8646502What's wrong with that?
>hgg slowly turning back into hdg
>general merger soon
uh oh
>>8646502Based af
Hail /hgg/
im going to blogpost the process of baking a lora and then not share it btw
>i'll go from not contributing to not contributing
riveting 2025 4chan discourse
>>8646513Obvious baitposting aside
I like that this place has very little gens, mostly just discussions
>>8646515Based? Why would I want your shitty fish when you've taught me to catch my own?
>>8646520kek
>>8646521I blame all of you for this
Sirs, a second kino has hit the general. https://files.catbox.moe/e04mbe.png
>>8646566Well, okay... I'll say this: I really like how the clothes look.
>>8646566can you post the dataset so I can check my config? I'm starting to think the old rex/came config was better than adamw.
>>8646566Is that the cripple girl from the cripple girl (video game)?
Based
low twintails, glasses, sex
>>8646583Syure https://litter.catbox.moe/3n7qeehe2au7q32e.zip
>>8646583cripple girl (cripplegirls)
>>8646599thanks. Painterly styles are excellent for testing whether or not the lora is learning all the details.
>>8646566.safetensors onegai
>>8646604is gyetting resized
baked a lora with a toml from this thread and it was SHIT
>>8646609Based. Return to hdg when?
>Gen image
>Damn this looks cool, I'll use hires fix to add some details to it, hopefully it'll keep the same image haha
>It changes the entire image
Why does hires fix (and img2img/inpainting) doesn't work for 291h bros
>>8645549it seems the infill/img2img setting no longer matters, and just img2img at 1.0 denoise will give the traditional infill results. Works pretty good.
>>8646618It does worse than base vpred?
bouguereau, oil painting \(medium\), traditional media https://files.catbox.moe/d4q88y.safetensors
rubens, oil painting \(medium\), traditional media https://files.catbox.moe/zyv14l.safetensors
himurohiromu https://files.catbox.moe/je01sr.safetensors
we high culture in this bitch lol
https://www.youtube.com/watch?v=OixAlpfEmfk
oh yeah van gogh and the guys artist tomorrow
although i'm open to more wacky classical guy suggestions
>>8646631any reason you use a trigger word in your style loras?
>>8646638it makes them better
>>8646618before trying to do anything else, keep whatever settings you have, add the 1536x lora at the end of your prompt on the hires step and change your denoise to 0.5
>>8646618>hiresfix>not cn tileOh nyo nyo nyooooo~. You only have yourself to byame, 291h cyad.
>>8646631Damn awesome
>himurohiromuIsn't that guy literally just copying Sugimori Ken?
What's the reasonale?
>>8646648i dunnyo i liked a few of his pics and it was 10 mins to bake
i enjoy baking :)
>>8646651post config then
>>8646630I wouldn't know, never used base vpred
Certainly does worse than illustrious mixes I've used
>>8646641>1536x loraWhat's that
Also the denoise is already at 0.5
>>8646647>cn tileWhat's that
>>8646631Not bad, non of these are painterly tho-ACK
>>8646658>What's thatA lora that lets you gen on higher base resolution but I found that it also helps on the hires pass, search for it, a kind anon share it some threads ago
>>8646631Why go through the effort of baking a lora when the end result looks like a generic realism slopmix
>>8646655Yeah that config is shit as
>>8646660 said. It gets close but doesn't fully grasp the brush strokes in the painting. Anyway we'll see.
>>8646663>>8646651>i enjoy baking :)
most disgusting bait I've seen all year but he falls for it anyway
>>8646658>Also the denoise is already at 0.5That's too high for hiresfix in my experience. Normally 10 - 15 steps and 0.2 - 0.3 will upscale without changing your base image much. Use 4xNomos8kDAT for extra clean.
>cn tileYou got a few options.
https://civitai.com/models/929685?modelVersionId=1239319
https://civitai.com/models/699930/xinsir-sontrolnet-tile-sdxl-10
>nyo you must be sardonic and angie all the time!
nyo! enjoy mozart! https://www.youtube.com/watch?v=-nG_YIDZpjw
>>8646660Low effort bait kek
>>8646631Damn that Rubens lora is pure kino, as was the van gogh
I'll test the Bouguereau in a minute
>>8646662I'll look for it
>>8646668Is it just an inherent problem with noobai? This was a non-issue with illustrious, like ntrmix or wai nsfw
Also what is that cn tile, you people use controlnet to upscale? wth
>>8646682It's always funny seeing it try to reconcile animu and the style kek
Also at least for me 291h was just shitty and essentially the only model I tested (today) that didn't do the character properly, might just not be stable enough for high denoise
>>8646668Lowering it to 10 - 15 steps and 0.2 - 0.3 denoise will just make the image blurry instead of adding detail, case in point
>>8646682>>8646692KEK they cast Elijah Wood to play N
Hmmm considering that I enjoy more painterly styles (but also overall style/prompt adherence), what would you recommend then?
>>8646682>Is it just an inherent problem with noobai? This was a non-issue with illustrious, like ntrmix or wai nsfwMight be the vpred. Everything you listed is eps or diluted enough to might as well be eps.
>cn tileYeah unlike the gacha roll that is hiresfix, with cn you can set the starting/end steps and weight of the model's impact on your image. Meaning, you can crank denoise super high for extra detail (not really my thing for henti) without having as much impact on your original pic, as your example showed.
>>8646693>Lowering it to 10 - 15 steps and 0.2 - 0.3 denoise will just make the image blurry instead of adding detailShouldn't be. What upscaler are you using? Show your metadata, coward.
Ouh, maybe William Blake...
>>8646693This stuff I genned with 102d custom
I think that people ought to bake for styles they want anyway, even inherent artists are kinda shitty
>>8646695I think it might be vpred
I have live preview on and what I see happening is that the first step denoises all into a blurry mess for no good reason like picrel
When I used those other models, most of the image was preserved and it only refined the detail which is why it gave me sharp results, I never even thought of hiresfix as a gacha roll because of that
>>8646697This happens with every upscaler, on that one I was using none
Here's one with R-ESRGAN 4x+
>>8646708>>8646711post your full hires settings you dummy dumb
>>8646715I dunno what retarded shit pancho's reforge setup from months ago included or not, compared to your hires, but I take it that 0 in cfg scale defaults it to the same cfg you used for the base gen, right? I ask because I have to slide mine over to 4 or 5. Well, back when I still used hiresfix.
>>8646720>but I take it that 0 in cfg scale defaults it to the same cfg you used for the base gen, right?Yes
>back when I still used hiresfixMoved to cn tile?
>>8646640you were right
added a trigger tag to my dataset and suddenly its a lot stronger
Why cant you train atop a shitmix? Weights are weights.
>>8645549>>8646628>it seems the infill/img2img setting no longer mattersthis is not true, check your metadata. it used to be that with the extension infill requests would disregard the denoise slider value and always use 0.5, while img2img would actually use it. but you can infill with varying strength now (try it on the website, infilling with the strength slider at 0.5 is completely different from infilling with the slider at 1).
"NativeInfillingRequest" in metadata = you used the "Inpaint Image" menu on the website/inpainted using "Inpaint Model" with the extension
"Img2ImgRequest" in metadata = you used the pencil menu on the website/inpainted using "Img2Img" with the extension
Both can have varying strength, and that is set with the denoise slider on the webui
damn, 5080S looks tempting...
He who shall not be named.
n-noodlekek??? leleck!??!?
Anyone got the link/torrent for v-pred 1+29 v2? Seems like all the 4chan archive sites broke their search, filtering for /hdg/ doesn't do anything.
>>8646927Some you can, people baked on AOM for ages
But they're usually malformed
without highlight faggotcocksucker bitch the hdg cattle are lost
>>8647046is this some nvidia guerrilla campaign
you said the same shit in hdg
>>8647100>you said the same shit in hdgHaven't been there for a month.
Just heard the leaks.
5
md5: 9d7d4615b8d27c6ad4b6d885e4038dd6
๐
you could not use the ellipses at least
>>8646927For whatever reason something bad always happens. If it works for you then congratulations but it usually doesn't.
>>8646927I'm not a ML expert, but I think it's because most slopmixes are fundamentally broken to begin with due to nobody having any idea of what they are doing and trying to patch more stuff on top just makes them fall apart.
>>8647172LoRAs are also just low rank adapters, they were never meant to be frankensteined into models.
Were we wrong about ayaGOD this entire time, hd- hgg?
https://civitai.com/models/1452858/slopslayer-lora-shiny-horrible-average-ai-style-to-put-in-negatives
You will not discredit his work around me, anon
>>8647235I tested it on release months ago, check the archive for XYs. It has heavy side effects on style and composition, whether you put it into positives or negatives.
>>8647235The thing is slop shiny doesn't even exist unless you use a shitty model with civit loras so why even use this instead of a good model
Lumina is flawed. It's just a bit bigger than SDXL but waaaay slower, and this is because of rmsnorm abuse. Just so you know it.
>>8647263thank you for your service
>>8647263This is what happens when you use a meme transformer and LLM TE.
>>8647278Still it's a good thing someone tried it, at least we know.
>>8647263lumina uses flux vae so obviously its gonna be slower since the latents are larger than sdxl
>>8647337vae encode/decode time should be negligible, it's the transformer model that just slower than batshit old U-net.
>>8647341sdxl has
4x128x128 for 3x1024x1024 image res
lumina/flux have
16x128x128 for 3x1024x1024 image res
can u see the difference?
the diffusion model is working on a much larger resolution
>>8647263Does it at least do anything notable
I tried that Flux Kontext and it was so fucking shit lmao
from what i see xerbatt is pretty damn good natively
anon, were you just looking for a stabilization here? i'm in the midst of looking at epochs
>>8647387genuinely might be the least change from inbuilt ivse seen in the 70 artists i baked lmao
van gogh, oil painting \(medium\), traditional media https://files.catbox.moe/f4sfhx.safetensors
xerbatt https://files.catbox.moe/y7ao87.safetensors
>>8647387Or using a shitmerge where he's much weaker
If anyone enjoys making loras, could I humbly request a Toono Esuke style lora?
>>8647410Very based van gogh lora
bros we ain't getting into AARU for gooning to mummy portraits!!!
fayum portrait, traditional media https://files.catbox.moe/ttmtcj.safetensors
>>8647433Not baking, but I checked his pixiv and he only has 7 images in full color.
>>8646031Seems strange complaining about this on the hentai board. Shame the /aco/ general is so useless when it comes to tech.
>>8647410You still up to classical artists? Here's a few suggestions if you decide to make another one
>John Everett Millaishttps://www.useum.org/artist/John-Everett-Millais
>John Singer Sargeanthttps://www.useum.org/artist/John-Singer-Sargent
>Gaston Bussiรจrehttps://www.useum.org/artist/Gaston-Bussiere
>>8647472Goddamn
Don't you tire of being based? Has the burden of such basedness not yet set you adrift?
>>8647475Yeah it probably wouldn't do very well anyway
>>8647478can't say i like bussiere and he has little on wikimedia but i'll bake the other two
if you scroll through 25k images of 1girl standing of course youre not gonna find shit you like lmao
are you people always this slow
>>8647410Thank you, it looks very sugoi on his own but I'll probably mix it with other artists since it does very ugly looking dicks lmao
Pic not related
>>8647487You seen above? 102dcustom got 99.5% on the way to him without the lora.
>>8647488yeah but I am not using 102d custom at the moment and yes, I tried his own tag before but I didn't quite like it
>>8647491ah
well good, it wasn't wasted time at least, enjoy it
gotta find some historical portrait artists to bake later
>few roman paintings survived antiquity
truly over
>>8647484I wouldn't expect you to bake any of them as this wasn't a request, just suggestions
So thank you
And also, the Bouguereau lora was really good, I actually found that the images you've made posted made it look considerably less accurate to the style than the model actually is
How do you guys process variation sets? Manually picking different pics or use some sort of perceptual hashing and setting some threshold for maximal similarity?
>>864751816 hash at 40 strength gets rid of all of them
i don't keep em, very few artists where i couldn't get at least 50 images
>>8647523oh yeah
in czkawka
Recommendation bros
What's a good epsilon base model?
>>8647523Huh, never used that tool, will give it a try. Thanks.
>>8647513might just depend on the shitmix
>>8647513millais https://files.catbox.moe/q6gt07.safetensors
sargent https://files.catbox.moe/v5i87i.safetensors
>>8647568Nice. I loved wd1.2 because it did classical styles so well but had an anime influence. For pure aesthetics I still haven't seen a model I like more.
>>8647568Awesome, I'll test it in a bit
>>8647586wd1.2 is like a 2022 model you don't want to use it lol
>>8647586What the other guy said. 24 fingers was the norm back then.
>>8646184with a dataset like that I feel like you're gonna have to make a shitty overfit lora on the few images that aren't outliers and use it to img2img and gen some new stuff and use whatever comes out well enough for a new dataset
>>8646184i'll try to bake it out of curiosity but what's your problems with it?
shit style or just the hands?
>>8646184I don't know this guy so you oughta judge yourself but here https://files.catbox.moe/lprisd.safetensors https://files.catbox.moe/7c61jk.png
although there is some fuckery with the hands, that might just be the dude's style idk
Is IOPaint good for cleaning up text in manga for datasets? Or is there something else? Or even better, is it unnecessary?
>>8647645novelai director tools
>>8647608why do so many use WAI?
>>8647645I've baked on fully paneled and text filed manga and got only minor leakage
I wouldn't bother
>>8647647its coherent with a wide variety of prompts, at the cost of some prompt adherence
>>8647648gotcha
I guess I'll just make sure to tag them with text
>completely start hating a style i had a phase for for like a week
many such cases
>>8647694I change the style of what I gen after like every coom
>>8647694this is why i can never grift
>>8647694This is why I gen all my waifus in their own unique style.
>>8647598Probably, his output is so inconsistent over the years besides doujinshi, but I don't have high hopes for a Lora trained on only doujinshi pages.
>>8647606Either the style is not baked in, and you get some shiny slop-style that bleeds through instead, or you get the style but anatomy becomes a mess (hands, legs, feet, etc). Basically it either doesn't learn, or overfits as far as I can tell.
>>8647642Thanks for re-baking it, honestly it works better than all my attempt so far so I'd love to know the config to see where I fucked up. I guess anatomy can still be a little wonky at times, but way better than on my bakes.
https://files.catbox.moe/ep8xj0.png
>>8647703I do this for only 2 of them
>>8647647What this anon said
>>8647649 is correct
Not just WAI imo, there are quite a few models that are very stable and easy to use in exchange for prompt adherence and lora compatibility (sometimes). Much to my chagrin, they basically always add in the official slopโข style lol
I wonder if it's possible to merge one of these but using one of the classical painting loras instead
>>8647734>>8639789I did retag it too. You seem to have used really low confidence? I use eva (goat) 0.35 with varied treshold (but 0.01 in this instance). I didn't even bother cleaning anything at all for this lol.
>did some yandere_trance with random characters
>coomed harder than with hard sex
remember kids, sex is the lowest form of erotica
>>8647766OH and did you name the directory 2_ for the repeats or just out of habit? I used flat 1 for everything.
>>8647768What a weird coincidence. I was just looking at the yandere trance tag on danbooru for reference yesterday.
>>8647767Thanks for the config.
I think I used wd-swinv2-tagger-v3, but then I manually went over every image in the dataset. Starting to think manually editing tags is a mistake, even if the autotaggers sometimes tags stuff incorrectly.
>>8647781i was randomly talking to my family about a minor obscure characteristic of a tomato plant and i got someone posting about it a couple of hours later on out so it is how it is
>>8647792All of these older taggers are pretty garbage compared to EVA, I recommend switching.
As for manual editing, I usually only remove random bullshit tags like alternate x and style parody and such with the dataset tag editor extension. And consolidate characteristic tags if I baking a chara.
>>8647793It's like we're all connected...
