Deep friendship edition
Previous Thread: >>
>>8624386>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
what's this https://huggingface.co/CreeperMZ/soft_noob_v/tree/main
>>8630793 (OP)>yuri gen for h opamazing idea op
file
md5: ff21ac6f13f0cebbf38cddf4e2f2fbe3
🔍
>>8630805No idea
>https://huggingface.co/CreeperMZ/soft_noob
>>8630814It IS /h/, is a mutual double dildo fellatio, my lawyer told me it was within the /h/ regulations
I wish there was a better way to do porn generations than using models based on danbooru tags.
>>8630820>porn generations>>>/aco/
>>8630820Wish granted, noob vpred also knows e621 tags
>>8630821Keep your autism in check please.
>>8630822I am well aware, but it is still imperfect.
>>8630826what are you trying to create?
>>8630826keep your porn out of /h/ please
>>8630829Smaller male on larger female. I have achieved it, but my prompt is ridiculously complex and easy to break. Takes a lot of trial and error to get a reliable and consistent prompt.
>>8630820What's wrong with danbooru tags?
>>8630841Yes this exactly, but for me it was clothing. Good luck trying to describe what a shrug is when you could just say shrug (clothing) and be on your way.
>>8630841To be fair, the cup size thing is exactly what we need on danbooru. Small/medium/large is the thing that is vague and different to different people.
>>8630858danbooru sizes are pretty well defined, people just don't understand them
>flat chest: A female character with a completely flat chest, that is, no breasts whatsoever. For breasts larger than flat yet smaller than normal, use the small breasts tag.>small: Small breasts are larger than a flat chest and smaller than medium breasts. About a palmful in size. In real-world sizes, this can run from just over AA cup to just under C cup.>medium: Breasts that are larger than small breasts and smaller than large breasts. Each breast is larger than a palmful, but smaller than the character's face.>large: Breasts that are larger than medium breasts but smaller than huge breasts. Around/more than the same volume as a sphere with the same diameter as the character's face, but smaller than their head.>huge: Breasts that are larger than large breasts but smaller than gigantic breasts. As large as/larger than the character's head but less than twice the size of their head.>gigantic: Breasts that are larger than huge breasts. More than twice the size of the head.
I just wish there was a way to prompt the model so that it understood that it shouldn't apply the tags to every part of the image and shouldn't mix characters up. It handles 2 subjects fine most of the time but if you do 3 or more then it really starts fucking up unless you use snakeoils or other tricks.
>>8630865The fact that we can even do two subjects pretty well now is a huge accomplishment in a year. Just give it time. Look at sora. Chroma will save us bwo.
>>8630865For me? Blushing rape ojisans. I find it cute that he's shy during the process.
>>8630841I understand this, I just wish there was something better anyways. I realize tags are the best we have right now. And as far as tagging systems go, Rule34 xxx tags would be ideal for me over danbooru or e621 tags.
>>8630854>What's wrong with danbooru tags?I don't do 1girl pinups or pov vanilla so they are woefully lacking to achieve some of the more precise control I look for in the sex scenes I do. Especially with positioning. I have to literally negpip some sex positions away to prevent them from being defaulted to.
>>8630865I know that pain. It was way worse in Pony though but it is still not perfect with Noobai vpred. I use the [SEP] function in adetailer to try and make differentiated facial expressions.
>>8630793 (OP)What's the style lora for this
>>8630862Also "hyper" at the end of the scale, from e621. Overlaps with gigantic for the most part, but using both works better than just putting emphasis on either one.
>>8630872>I don't do 1girl pinups or pov vanillaI do and I still find the existing tag capabilities lacking.
00200
md5: 16aa98c2380e0c4431fa45202d1d99a0
🔍
00201
md5: fefc541ddfd11faa8a5c1fb2da2785e5
🔍
sisters?
how good is 7s/it for bs3+gradient+memefficient?
and how to prevent te training in kohya gui?
with --unetonly it still says te modules: number at the beginning of the training
>>8628917>how much GA?1
why is there two generals?
>>8630948wait a few hours and you'll see
00203
md5: ad1e097c76082175a8284c547bc7f72b
🔍
>>8630948Is this our first /hgg/ meme?
What's the difference between this and /hdg/?
>>8630971Nice thighs. Catbox?
>tfw can't maintain your usual style
What are some heavily biased round anime face artist in most illustrious checkpoints?
>>8631027It's mainly nyanpyoun
>>8631069>nulloh he does draw nice thighs, thanks!
>>8631051is your "usual style" 20 layers of random noise and gaussian blur?
>>8631051If you're going to post here stop using a name, otherwise go back to either /hdg/ or /b/ please
>>8631106why would you complain about a free filter
>>8631076I don't like the digital look.
>>8631106No.
Sometimes it's worth taking the overtime.
https://files.catbox.moe/jneg4k.png
Does anyone know what I'm doing wrong here? I have my GPU weights set to 256 if that matters at all.
>>8631156>latent>0.3 denoiseYou need like 0.6 to make latent work.
>>8631162I'll try that, should I be using anything that isn't latent?
And does GPU weight even affect the quality at all?
Animation gens better off posted in >
>>8627272
>>8631186I dunno, depends on what you want. I personally don't think latent+high denoise is worth the headache.
>GPU weightI don't know wtf are those.
>>8631186If GPU weight is how much of the checkpoint is loaded into GPU VRAM and how much to RAM, then it doesn't affect quality at all, only speed.
>>8631142xir, how do these knickers work?
>I don't like digital style
>so I make my gens look like a blurry photo taken with a 2005 digital camera that has been compressed a million times
I swear to God, these people have never seen a painting irl
>>8631208I don't think he necessarily aims for paintings?
I'm not sure whether I like it or not but it's certainly much more interesting to look at compared to usual waislop so if he can't make his gens look less ai by other means, it's whatever.
if your gen has a background it's slop
>>8631210But if you look underneath all the smear and noise it's WAIslop down to the pure white highlights
>>8631189what is else to do here then?
>>8631219uhm.. sweety? We discuss the intricacies of lora baking and traps (formerly otoko no ko).
The loras affect both the woman and the mans face in my gens when i only want it to affect the woman. How can i fix this?
>>8631245Do not use jeetvitai garbage
>>8631286unless adetailer is the cause
>>8631294don't run adetailer on the men then
>ADetailer
What year is this
>>8631297Are you trying to imply something? Just come out and say it.
>>8631298that is a shitpost, although I personally don't use adetailer
I either don't care about a gen and throw it out, or I do care and want to improve it and then I'd rather upscale and manually inpaint faces to cherrypick the expression
>>8631298I'm implying that you should inpaint what needs to be inpainted rather than blindly letting a model trained on photos of real humans decide what needs to be inpainted with settings that you have to set before the whole gen is done.
>>8631309Yeah, that's fair. Other than the real humans bit, you can use a different detection model if it ever becomes a problem.
>>8631305Same, but lately I've realized I tend to use only like 0.2 denoise anyway, not enough to change the expression.
where can i find a good lora for ryona?
Been away from genning for a few months. What's the UI for the cool kids right now? I was using ReForge before
>>8631427we are still using reforge
>reverse image search
>zerochan
who uses this awful website?
>>8631428speak for yourself -- reforge isn't even updated anymore
>>8631436I know pancho leave us some time ago but I still use reforge and don't find any reason to stop using it
>>8631436>reforge isn't even updated anymoreDo its features degrade with time without updoots?
>>8631433now that's a relic from the past
>>8631420anyone? civitai has nothing I can use
>>8631659not the right place to ask, try /d/
>>8631164These really perplex me. Why does the dude animate the most atrocious sd 1.5 slop? we're talking one second on a 4090 for the gen and five minutes for the animation
>>8631420why need a lora for it unironically
lel
md5: 61d4bd4c6b85569310685414418aece5
🔍
grorious
>>8631847i would share this but it's 48gb uncompressed
hm
how do you get the seed from a preview pic you generated on comfy
>>8631140why is comfy so disgusting to look at?
>>8631294bwo where's your [SEP]?
>>8631949Explain it to me, my man. I normally inpaint when I have more than two faces on screen.
>>8631976>>8623505why do we have two threads again
why does resizing a lora that's used as base weights have such an arbitrary effect on vram consumption? with a sv_fro resize at 0.95 training hits 10gb of vram consumption. at 0.955 it uses 9.
>>8631976https://github.com/Bing-su/adetailer/wiki/Advanced
have anyone tried light2x lora?
It's fast but I feels like it doesn't work well with current wan's loras.
With light2x, I get like 30-40% other lora movement compare to just standard wan+lora
There was some discussion recently about banding and I suppose potentially other artifacts. Well these are issues that the ML community has tried tackling even before diffusion models, and as it turns out, there are dedicated debanding models as well as other deartifacting models. There's even halftone removers. Although they are imperfect. Here is a debander.
https://openmodeldb.info/models/1x-Bandage-Smooth
I just tried this with Noob vpred. What I can say is that it does seem to mostly work, but sometimes Noob's "banding" is just categorically a different kind so it's not entirely effective depending on the seed. The model also sometimes confuses some barely visible lines in the image for banding and smooths them over. Not surprising. For instance, if I prompt for a wood tile floor, some of the faint lines in the wood are smoothed out and even less visible in the filtered image, so the image looks less detailed on some parts of the wood. Honestly though it's not a big difference and the effect on banding is more noticeable. I think this might be a good default filter to have for art styles with flat shading. The model keeps strong linework untouched while generally cleaning artifacts, so it should be pretty effective for just improving Noob outputs depending on the style and whether or not you're using a stable lora.
Hey guys, haven't been here in a long time. What is currently considered to be the best models for image generation? Is it still pony variants, e.g. autismmix?
>>8632084After I am done with my current backlog I am going to tests out this one, perhaps It would the last and final step I was looking for on my current workflow
>>8630793 (OP)for a tech illiterate should I go with reforge or comfy ui?
>>8632128reforge is the most out of the box experience so yeah
>>8631140Can you share your Workflow, please?
I want to also check out that node.
>>8630793 (OP)I just saw that ReForge is dead. Should I just move to Comfy?
>>8632187Why do you think moving to comfy is a good idea? What is comfy going to do for (You) that reforge isn't already doing?
Constant Lr 0.00016 with AdamW, BS 3, dataset is 22+7 dupllicates for one character, 16/8 dim/alpha, 3000 steps, but the output still feels underbaked, prompting without details works 50/50. Training and testing on the same mix.
What am I doing wrong?
>>8632196have you tested any prompts from the dataset?
>>8632196Share your full json but
>constant>0.5 alpha/dim>training in shitmixesAlso a lot depends on tagging
I need to know loras and bases for this, HELP PLEASE
How do you get NSFW shit to actually work with Wan? Like it's trying, but it's all stiff and slow, even when prompting for fast motion and stuff. I turned down the weight of lightx2v's lora and I'm using a NSFW one, but it's still kinda not there.
>>8632194Eh, mostly to keep stuff updated. I already had it installed from when I tried it earlier this year. I don't mind the node autism desu.
lil bro needs advice on wholly personal decisions he already has opinions on
>>8632269really shitty bait
kill yourself
>a few artists with the same random guy collage as top image
>turns out it has like 500 artists tagged and it's high scored because of meme or whatever
lol
23
md5: cf969294551e71a7bf098d9d2b9159b2
🔍
>>8632217Reference pic related.
Yes. With an exact prompt from the dataset it starts to work fine between 600-1200 steps already:
https://files.catbox.moe/vyyid2.jpg
https://files.catbox.moe/psnb7a.jpg - specifying "blue bodystocking" bleeds blue on the turtleneck (may be the problem of the dataset), omiting "thigh boots" makes them regular height.
https://files.catbox.moe/oehri1.jpg - adding styles worsens the problem above.
https://files.catbox.moe/ln2lwj.jpg - almost exact prompt from datased, but for a sample not present in it.
https://files.catbox.moe/4397p6.jpg - almost empty prompt.
>>8632230https://files.catbox.moe/j5bjit.json
>constantI was trying to fry, that's why lr is so high also.
>0.5 alpha/dim>training in shitmixesWhat's wrong?
>Also a lot depends on taggingWd-tagger + manual postprocessing. But I was not sure should I specify color for bodystocking, because there is no such tag on danbooru (and didn't).
Sorry for not /h/ (and even /e/).
>>8632338lower alpha just lowers your lr in effect
also shitmixes are fucked up for training, the "train on what you gen with" is retarded advice
anyway your conv dimension are low as fuck, i personally had to keep them at the default 32/16 even when changing the real dims if i wanted anything usable
>>8632350> lower alpha just lowers your lr in effectBut I've seen 0.5 in multiple configs here.
> also shitmixes are fucked up for training, the "train on what you gen with" is retarded adviceWhy so?
> nyway your conv dimension are low as fuck, i personally had to keep them at the default 32/16 even when changing the real dims if i wanted anything usableI can't even change them in kohya gui. Why not 32/32?
Anyway thanks, I'll try 1 ratio for network alpha, Illu as the base and 32/16 for conv.
>>8632359>Why so?Not him but may I ask why you think it's good to train on the mix? Besides being the closest possible to the dataset.
>>8632359>But I've seen 0.5 in multiple configs here.I mean yeah but if you're talking not baking properly that's one thing to look at
>Why so?They're not clean, it's just a bunch of misaligned layers with fucked up values that fry under training.
>I can't even change them in kohya gui. Why not 32/32?That's just what the default in ezscripts was, I modified it both ways and it was just worse so I left it as is.
Kinda interesting thing: the top scored images on danbooru are like 70% westslop
But the top rated images from the top image count artists are almost exclusively anime
>>8632367Because the model changes the weights and observes the effect during the training. Basically bespoke.
>>8632370> They're not clean, it's just a bunch of misaligned layers with fucked up values that fry under training.But training algorithm should see when it's frying more or getting closer to destination.
>>8632405Well in my experience things were always "weird" when I trained on the mix, both on pony and now on noob. It would learn like 80% of the style, then take a sharp detour into strange territory that's unrecognizable from the dataset. Training on the base model was always better.
It's even worse when you consider that 99% of slopmixes are awful mixes of eps and vpred so you're just violating the poor noise at that point if you bake on it.
Luckily illustriousXLPersonalMerge_v30Noob10based.safetensors is a good mix.
Luckily [HEADCANNON] is a good mix
What's the reason Illustrious doesn't recognize so many characters and artists with tons of art & why hasn't it improved in any way with subsequent releases? Was watching a friend use NovelAI and it recognized near everything.
>>8632445Use noob instead?
>>8632448Not really seeing a difference in my experience, it knows and doesn't know the same characters and artists
whenever I go back to adding Lora's it gives really unstable results
>>8632445Name these artists and characters.
>>8632464Countless, but Yoshitaka Amano as an example for an artist it does't even try to emulate. None of the characters from Unicorn Overlord work either
>>8632454cry about it, most my stuff isn't suited for this site so I picked something random
>>8632468your bait is boring ass
>>8632472yes he is
he is sperging out in hdg too
>>8632468Are you retarded or just baiting?
>prompt for "suspenders hanging"
>just gives me normal suspenders
...
>3250 artists got into the second round for testing
whew
>>8632495what are you testing lil bud
>>8632502i'm just looking at artists to bake loras of
first round was 20-something k of just top scored pics
now im doing 2nd, 3rd top scored and latest pic of each of the 3k
also god bless chatgpt for optimizing its code by a factor of 25 just because i asked it to
>>8632503could you add Nanameda Kei to the list?
43535
md5: 35a65bf14bf096335abc1a6d13476bcb
🔍
Lmao this is funny
Maybe I will
test
md5: b0c1be5caf8f87280c27685f989a818b
🔍
>>8632453>>8632464nta but is that about Illustrious 2.0 or can you get the later ones somewhere? I tried a few literally who artists on that and they were all more accurate on base noob v-pred.
>>8632509What was your original prompt? Make me a bot that scrapes danbooru?
>>8632529A modification for my first script but yeah
>>8632534Why aren't you using grabber?
>>8632536You say this like there's a comment chain I'm butting in on rather than you talking to yourself and posting pics over the course of several days. Doesn't really matter that much anyway.
>>8632537tldr im not download some one artist but a bunch of posts from a bunch of them and i want them named a specific way and then i want to use an artist list blablabla
>think about getting into training
>check the OP
>see big rentry
>stop thinking about it
>>8632542Should just ask for help then. You should be specific with your questions, however.
>>8632542just make sure never to listen to lora training advice and you'll be good
t. gives lore training advice
>>8632084This is not doing anything for any of my gens that clearly have that banding effect on them
only dataset matters
but also this
>>8632546
>>8632542it's outdated anyway
>>8632543Well ok.
Is there some one click program for training a lora? Assuming you have a dataset already prepared. I see the easy training scripts github but somehow I feel like the out of the box settings it comes with is likely garbage. I just simply don't want to mess with adjusting settings or to spend time knowing what everything is or does.
>>8632556You can import someone else's config using easy-script and just click run.
>>8632562Is the easy script what everyone is actually using?
>>8632566that, kohya and I think there was one more
Doesn't really matter, they're pretty much just UIs/CLIs on top of the same thing.
>>8632566That, SD scripts, and kohya-ss. Should also be able to search configs in the archive with "config." There were a few posted recently.
>one artist is 200 same look pics of the same character just in different poses
huh
weird hobbies people have
>>8632651One time long ago there was such a thing called a waifu.
>>8632652I think that's a mental break because he had a couple of pics of other characters before before over the years transitioning into doing just that one for years
And the exact same style every time
Almost makes me thing it's AI
Let's say that you want to train a lora to be able to do a small thing like add a piece of clothing. In theory would using a database of images generated by the model but inpainted to contain that piece of clothing be a better way to make the model learn without affect style or learning other weird things?
>>8632549Like literally no effect or just very slight? It's definitely doing something for me, but the amount seems to depend on the image.
>>8632662Little to no effect, perhaps I am doing something wrong, should I use it as any other regular upscaler on the extra tab on reforge? (even if I am not upscaling)
>>8632654I asked much the same thing a month or two ago. Anon said the seed you set during training doesn't work the same way as the seed used when genning. If true, that means you'll still get some loss from the whole pic. That brings with it all the usual issues of training on AI, mainly a bias towards broken anatomy even if the anatomy in those dataset images was totally fine.
Same anon suggested a different approach I've already forgotten the name of. One used specifically to train on image pairs, focused on learning the differences between them.
Haven't done anything in a few years, what are some good uncensored models for anime style?
>>8632721start with noobAI or one of its merges
I'm looking for the guy who made the Fua1heyvot4ifsr lora, just wanted to say thank you, and also wondering which model you've used for the pic you posted in the previous thread (box would be nice :P)
I finally thought about making a tag list for filtering out of datasets
I usually only clean datasets of garbage tags like alternate tag and the character tags anyway lately
>30500 total images to bake
i will be done next year
>>8632654>>8632720https://rentry.org/Copier_Method_Notes
This is the type of stuff that should be in the OP but only gets mentioned whenever someone wants to "win" an argument against a newfag.
well if there's one thing that's nice about scraping an artist from only danbooru is there's almost no garbage to clean up
never scrape r34 bros...
Anyone know if there's a way to negate wrinkles on skin tight and impossible clothing? I want to make it look like another layer of skin rather than clothing. I've already tried negpip and regular negatives with stuff like "taut clothes", "skin tight", "seams", "wrinkled fabric".
>>8632777if it's flat color just erase the lines yourself
>>8632777you could unironically try baking a lora
get a bunch of these pics and erase the wrinkles and bake
although i can check later if thats a problem on my end in the first place
could just be a setting issue
>>8632471legit where is the bait
go gen something in Yoshitaka Amano style without a lora and show me the result
>>8632800It's not :(
>>8632806Damn. I don't want to take the training pill...
To be fair, it seems like I can minimize the lines by using the latex tag. But they're still not completely gone.
>>8632811Why would a model trained on the danbooru dataset be able to replicate the style of an artist with less than 95 posts on danbooru? And why wouldn't I just get/train a lora if I cared about genning schizo watercolour porn in his style?
>>8632816Because it is just an example and I've seen it struggle with artists that have over 500 pieces while nailing artists with 50.
>>8632819That's what I'm genning right now, except they are peeing while being strangled and anally fingered. Very fun, try it.
>>8632816I thought he was talking about stuff that works on noob and not on illustrious?
>>8632818The more different a style is from the anime average, the faster it trains. Same goes for Loras btw.
>>8632819technically /d/ but we have so few gens here I'll let it slide
once again i am hit with the curse of more anime datasets not really wanting to bake properly
>>8632816The deeper question is how is Novel AI capable of doing that?
>>8632826Noob and Novel AI.
>>8632819Hot, but yeah /d/. I really wish /d/ wasn't filled to the brim with futa so I could go there.
>>8632509https://en.wikipedia.org/wiki/Recommender_system
>>8632769>https://rentry.org/Copier_Method_Notesi hate to do this but this definitely deserves a mention in the op
>>8632871illustrious and novelAI
That's what confused me, he should be comparing NAI to latest noob v-pred, not the underbaked model from 10+ months ago.
>>8632875Then put it in there when you bake
>>8632871>The deeper question is how is Novel AI capable of doing that?So you're asking how they trained the model to achieve that? They actually wrote about it in that shitty research paper they released sometime last year
poppowerhearts - https://files.catbox.moe/e0piil.safetensors
>>8630793 (OP)>AI is a slo-It took me 3 days to generate and inpaint (and photoshop) to create a [spoiler]single[/spoiler] good fappable image
file
md5: 6248286c45dcd8080142a5a26619d39d
🔍
>>86329193 days for this? I mean, looks fine but, 3 days really?
>>8632924i think he stole this from some /g/ post
>>8632925maybe, just maybe giving the namefag some attention for his gen progress video on the other thread was a mistake
>>8632912Could this be achieved with weighting on tags for loras?
>>8632925I thought the miku inpaint anon was from /hdg/ because we rarely see him so I guess not...
https://gelbooru.com/index.php?page=post&s=list&tags=princessofpower
Looks like a cool artist for lora except most images are about 500x500 kek.
If my my attempts to bake a lora were more successful I would bake a lora for this guy
>https://danbooru.donmai.us/posts?page=1&tags=xerbatt
>>8632538When? Now, anon!
>https://www.illustrious-xl.ai/
file
md5: 8f6aad65755e47637efab09c2a0806c3
🔍
>>8632974Uhhhhh where is it then
Speaking of bakers, the anon who mentioned making a Lune lora never came through? Guess I should try to find the less slopped one on jee-
>just bake it y-
Anon, I am retarded. I see your lora convos and it's nothing but hieroglyphics and runes.
>>8632981Uh.. right here!
>>8632974just look at those preview images!
>>8632940I have no idea, try it
What's the state of the art image gen model right now, for local gens? I've been out of the loop for a while. Last I was in the loop it was pony
>>8632974>>8632986T-the power of 3.5... wow...
>>8633016It's better than it was on the pony era
disembodied penis bros, what's the prompt to get the girl directly from behind? Pov keeps spawning a stump.
>>8633040have you tried "from behind" or "rear view"? maybe add "straight-on" and avoid using any tag related with the face or/and front parts of the body
>>8633044Yeah but it's always a three quarter view from behind. There's no real tag for 0 degrees from behind and I figured straight-on was a front tag only. Regardless I got it by lowering the image width.
>>8633045try "back" next time
>>8631947whats disgusting about that?
>>8632919I would be much faster to pre-draw stuff then do a low denoise inpaint over it, than to keep rerolling the same area for twenty minutes.
>>8633045you can also cheat with backboob
but from behind+straight on works fine for me
>>8633067Nice fishnets, mine always look like shit
>>8633080well these look like a skin texture not a piece of clothing
I think it's the lack of outline when everything else does have one
>>8633080Thanks
>>8633082Might be cause that character is supposed to have white pantyhose and fishnets but it missed the pantyhose part.
anybody figured out how to make a lora extract from rouwei to use it with noob model?
>>8632675Not sure about reforge but in theory it should work unless reforge does special things. In Comfy I do just load it like any other upscaler model. If you download a different 1x model from the same website (that has a noticeable effect), does it work?
Like this one https://openmodeldb.info/models/1x-NormalMapGenerator-CX-Lite
I haven't tested it though.
>>8633116Yup, that one works, so do others that I have like nomos8k and anifilm
if I use an art style XYZ Lora but use this prompt at XYZ:0, will it be the same as me not using the lora at all?
same for if I use it at XYZ:0.3, will it be too small that the end result won't look any different compared to me not using the Lora at all?
>>8633130just... do an x/y
>>8633009Expedition 33 waife
https://files.catbox.moe/je4xlp.png
>>8633130This type of thing varies per lora or per artist tag and you'll have to test it out yourself.
file
md5: 0d623429429c55fe794a55f31e730080
🔍
>>8633162baasssssssssssed
>Wonder what the latest best model is
>let's check /h/ to see what they are using!
>it's all dogshit
Oof. We are fucked, huh?
>he says while posting dogshit
>>8633188>too pussy to quote>too pussy to post theirs
>>8633186can you pinpoint said dogshit?
>>8633198can't really argue against that
>>8633195Your eyes break or something? Look at something like
>>8630970>>8631051>>8632850There's been 0 progress in the past six months or so
>>8633202I mean, it's not much different from what you posted so...
The tools are already there, it's up to the sloppers how they use them
Just by mere curiosity, what did you expect to see?
>>8633205Something vaguely decent. Not like it was made with 1.5
>it's all dogshit, not like my shitty sloppy gens
You're in the wrong thread, shitter schizo.
>>8633208We have those since the pony era bro, your bar is way too low to the point of being irrelevant
If you couldn't make anything good last year, you are not going to do anything good now
>using NAI
>important image settings
>"Model switched to anime mode"
Huh? There's not even a option to change anything like that
>>8633190Sorry my wife's pussy is for me only.
>>8633212>le smeary eyed gangbang with faceless men
did hdg get nuked again or what's up
>>8633217Are you really not going to try to one up my dogshit pony gen? All bark and no bite as always huh
>>8633225No one cares about matching your garbage because that's what it is, garbage. Cute that you desperately tried to make a new gen but no one has been impressed by group pictures in years.
why do we keep getting ""people"" like this here?
>>8633234Because the schizos in hdg finally found out everyone sane left and they've been circlejerking for weeks now. The only thing missing here is the Tifa spammer to complete the circle.
did anyone actually think a thread where half the posts are noodleschizo would have any future?
Anyone else using 4.5 have sometimes where you'll use a character tag and it seems just that character tag for some reason changes the style? I've been noticing this and it's pissing me off
Is right click to view metadata gone, or am I just stupid
>>8633208Your only criticism is about the style some people used for their gen, the fact that you think this is a good way to judge a model strength just show how much of a retard you are
>>8633279Yeahhhhhh, idk what's going on.
Can someone try galleon (granblue fantasy) and see if they notice a change? It's like...it just becomes shit but only with that character tag
>>8633279havent had this problem yet on 4.5 but its really bad on noob using any senran kagura character
>>8633293Try galleon (granblue fantasy) next time you mess with 4.5. It's doing the same shit and I don't know what it is. I thought maybe it was the granblue fantasy part, like it was reading it and using it as the series tag. But other Granblue character works fine
>>8633149OH that might've been me
I can try baking it but there's not much data for her desu
>>8632445>Was watching a friend use NovelAItf? who does this?
>>8633327Oh yeah it's like 99% 3d shit, I doubt I can bake anything usable from it.
>>8633329I imagine th fucker's like a kid watching his brother play CS1.6
>>8633332>it's like 99% 3d shit, I doubt I can bake anything usable from it.But bro here literally baked a lora from gopro images and now posts beautiful mushy painterly 2d anime with it?!~?!??!?!?!
>>8633214>outing urself as a furfaghaha
>>8633214It's a small icon in the top left of the prompt box.
anyway i'll bake it but 2/3 of the dataset is 3d, i tagged it to help offset it but i doubt it will look that good lol
the 12 2d images that are of her aren't really detailed enough by themselves to bake imho
>>8633341Honestly the ingame models are pretty good, I'd just use screenshots in different outfits and from different angles. It would be better though if someone ripped them so I could pose them in blender and not actually screenshot from ingame.
>>8633341>>8633344what are you talking about?
>>8633346Somebody requested Lune from Expedition 33 and some other anon is trying to bake her from fan art (which will probably not work out).
>>8633327>>8633332>>8633341>>8633344Lunefag here. I figured the source of the character would make it difficult. Especially as it's still a pretty new IP and everyone is rushing out to slop her in blender/3d environments. Probably why the non 3d stuff of her on civit looks awful. Even by civit standard.
Not to mention, all the "anime" fanart of her somehow manages to remove her asian aesthetic, which is her most defining character feature. Well, after her feet, sure. picrel is probably her most asian looking booru art. If you squint.
Reposting a question I had in the other thread: Anyone ITT make an Illustrious Hatomame (Ago) LORA? He's a pretty good JP artist and there was a Pony LORA a while back on Civitai, but none for any Illustrious version. If he was a normal artist I'd think about making one myself but all the stuff I can find has lots of dialogue/SFX so it'd be a pain to clean up a dataset. I checked the LORA links in the OP but didn't find anything, but I figured I'd ask in case anyone has a personal one.
Also, if anyone would be willing to make a dataset, I'd be happy to train the LORA myself.
>>8633202>get singled out by a guy posting barely /e/ shinyslopan abstract kind of feel
That... actually didn't really bake in much of a 3d style. Granted you will never ever get the tattoos right but it works and doesn't look as sloppy as the civit stuff?
I tagged it with lune \(clair obscur: expedition 33\) and tattoo https://files.catbox.moe/t9ijj5.safetensors
>>8633363Text and SFX probably wouldn't be that much of an issue on modern models, no?
>>8633377prompting for 3d and photorealistic also works lmao https://files.catbox.moe/z4v6r0.png
>>8632359> Anyway thanks, I'll try 1 ratio for network alpha, Illu as the base and 32/16 for conv.I tried this and lora baked out even worse. At least it can work with noobillvpred1.0, but with over saturated colors.
>>8633377As long as you tag the style of the dataset images, it shouldn't bake the style in too hard.
>>8633363>300 pics on Gelbooru>3 pics on Danbooruremind me again why people train on that shitty site
>>8633388high quality standards
>>8633377>Text and SFX probably wouldn't be that much of an issue on modern models, no?Hasn't been my experience, sadly. I'm sure it's fine if you only have a few images with it, but I've seen some Civit LORAs that are really lazily thrown together, and so they generate random speech bubbles, SFX, logos, etc. I don't mind a blurry patreon name here and there but there was one LORA I tried where I could identify the source images because they all had the same stylized CG set title burned into like 70% of generations. That said, I haven't trained a lot of Illustrious LORAs so if there's a way you can fix that sort of thing in training I'd be happy to gather the dataset myself.
>>8633391I think as long as it's tagged it should be okay? Civit bakers always fry the shit out of their loras and tag them with 2022 taggers so that's another thing.
>>8633363What about this one? https://civitai.com/models/1661404/ago810-style-noobai-illustrious
Also sadpanda has textless versions if you want to bake one yourself.
>>8633377>Text and SFX probably wouldn't be that much of an issue on modern models, no?They will start to show up if your dataset is full of it, even when properly tagged. Easy to remove with negs usually, but having to do that in the first place is annoying.
>>8633394Huh, I missed that one when searching "Hatomame", "Hatomame Ago" and "Ago", weird. Well thanks for that, I'll have to try it out.
>sadpanda has textless versionsI searched there and only found text versions outside of the game CG sets, which might end up with weird backgrounds or even as a character LORA since they tend to only have one or two main girls. There's some older doujins that are textless but I think I want to focus on his more recent stuff since he's really improved artistically in the last few years. Much more refined. Still, I guess I can combine it with the game CGs and some of the newer doujin pages, I'll just have to tag more aggressively if I do end up baking my own.
>>8633248>noodleschizowho? Oh that's you isn't it...
>>8633279I remember this being a thing back in the day on early NovelAIv3 but even on Noob Pokemon characters change the style (for the better).
file
md5: b35a659c09fa1d1a7a2f7131aa585971
🔍
>>8633289Not him, and you're not even wrong really, but aesthetics really is what most people care about at the end of the day. Composition is important and the cool stuff you achieved with controlnet matters, but as long as there aren't any glaring mistakes it's alright and people will be focused on the aesthetics. Inpainting 1 million times to draw Miku is only rewarding if (You) care that much not because there's actually a "correct" way to gen.
I have all my datasets ready, and I've already baked 12 artists so far. It's going good.
file
md5: dfe70eaa595213e0481ec70589b4022e
🔍
>>8633350I do a lot of vidya loras and you really only need a little 2d. I mean I made an Eve (Stellar Blade) lora with several outfits a month after it came out last year by lurking at taking people's screenshots and vids from /v/. At best like 10% of the large dataset was 2d and it still worked out. Admittedly things are a little different on noob than on pony though. Is there really 0% 2D for those characters?
>>8633407>which might end up with weird backgrounds or even as a character LORANah I've made loras from h games and it's fine if you tag all her features.
>try prompting for a relatively obscure position
>anatomy goes to shit
Sigh.
>>8633554thissssssssssssssssssssssssssssss
>>8633554Yeah but it's always preferable not to resort to that.
>>8633549Why don't you post it anyway? Maybe we can help.
>>8633558I am not joking
>>8633560true but
>relatively obscure positionhe knew what he is getting into
The schizos really can't survive here can they? I've never seen such a strong effect in any general. Like someone put bug spray in the air or something.
>>8633575>namefag came here>got made fun of and then ignored>left to cry in hdgYou love to see it
Do you guys have any tricks to more reliably generate complex scenes with two or more characters ? For example, one girl riding cowgirl while another masturbates in the background, etc.
I can manage more complex scenes maybe ~30% of the time with regional conditioning but it isn't very consistent. I don't really know how to prompt the model for scene composition
>>8633595the less painful way to do it is with controlnet either you find something very close to what you want to do or trace over a bunch of other images into a new one to get the composition you want
>>8633595pick up a FUCKING pencil chud.
>>8633599Will try
>>8633606I have been getting good results with inpainting except for fixing minor faults, but trying to get scene composition right with it hasn't worked for me
>>8633619I'm just trying to jack off here
>>8633450I would argue that the post he quoted have 3 really different style which is a strength but you can also train lora for style easily however model knowledge and anatomy is the main thing you want to judge since its not as easily fixeable
Alright I give up
What are the Pixiv Asians doing to get the smooth jiggly wan2.1 videos?
>>8630793 (OP)I've been trying to gen some girls holding giant syringe no needle full of cum
going for that (artificially) pumped full of cum vibe
all I've been getting is slop even when I inpaint and img2img hard, the model itself just doesn't get giant syringe full of cum, quick search on the boorus confirms there ain't shit for training data
I'm wondering if anyone got a better model, maybe with natural language processing that can gen it
final option is of course train a lora but then gotta make training data out of practically nothing and lots of it to make it not just one pose and flexible
>>8633674bwo is our word. you may use bwudda or bwuddie.
>>8633699https://litter.catbox.moe/zibz2b8l7z9uy1hr.png
doesn't seem like it'd be impossible to get something ok with inpainting
>>8633710I can get something of similar quality too
I'm trying to go beyond this to somehow get something more stable and detailed
like no needle, less bad geometry of the syringe, and of course as always, good hands
>>8633716nta but, remove the needle with ms paint or something like that and doodle around to help inpaint to fix the hands, you are not going to get all what you want with mere t2i
>>8633673>however model knowledge and anatomy is the main thing you want to judge since its not as easily fixeableYes I agree. People started memeing me when I told them object permanence was a far more important problem to fix than the sepia autism was known for. NAIv3 could do it pretty well but pony couldn't but the poor reception to that obvious fact is kinda my point. People are more willing to reroll bad anatomy or controlnet it or even draw it themselves to get that stuff apparently, but aesthetics are essential even if you can train styles or fix colors easily.
>>8633751I still don't get how sepia became a meme when it was only a thing if you ran no artist butful_score_9
probably nai shilling tbhdesu
>>8633758pony was perfect, butiful and valid, only a paid indian shill would think otherwise.
i forgot i bunch of artists i used to like :(
>>8633767actually fuck, is the old 1.5 lora git archived anywhere? i could just look there lmao
>>8633731Like cccino on Pixiv, don't want to upset the /h/ mods
>>8633800>don't want to upset the /h/ modsmods don't get mad, it's people ITT who do and report you
re:re:Re:Re:re:re:re stabilizing artists that already work
base kotteri really has problems with contrast (most stuff does on noob but his drawing style is even worse for it)
a lora definitely helps, and again i can see the linework just being sharper with the lora for whatever reason
i had my cfg at 1 this whole time ﹏
>>8633852strongest /hgg/ genner
I need your strongest config, lora baker
>>8633852Without a CFG++ sampler?
>>8633865i bake with https://files.catbox.moe/65wttm.toml
just remember to shuffle captions
>>8633865the glora one is pretty good
>>8633921I use bare sd-scripts but you should be able to port these to easy-scripts.
https://files.catbox.moe/6re9os.toml
Do you guys inpait all areas of a picture one by one, or is there some other way to get a pic all polished?
>>8633979I only inpaint the most important parts, backgrounds mostly remain untouched
>>8633979If something takes more than 3 inpaint passes then it gets abandoned
>>8633979I actually don't inpaint. I try to get a prompt as close as possible to generating the thing I want, get the closest cherry pick, then go into krita and draw over it a bit to fix things I want, and then go back and do an img2img.
>>8633979dyepends but i honestly havent done major inpainting since pony
>>8633979Umm sweaty? Just reroll. If I wanted to draw I would pick up a pencil?
am noob i want to make my own loras, is this a good guide to learn from?
https://www.youtube.com/watch?v=m3ENCAwWDXc
>>8634016No one here uses flux, saar
I'm looking at noobai controlnets right now. What's the difference between pidinet and hed? Which one should I use for transforming my doodles?
>>8634022how complex are your doodles?
>>8634026Pretty simple, like rough sketch type stuff. Are you saying complex ones need a different one too?
>>8634027Sometimes they do but if your doodles are on the stickman side use lineart_standard (from white bg & black line)
>>8634029Uh what if they have color and I want the color to guide the gen?
>>8634030you mean like they are sticks but they have different colours? if so then yes, use that one
>>8634032No they're actually roughly what I want but just badly drawn and simplistic. But if that one works with colors then I guess it's what I want still.
>>8634034You could also just don't select any preprocessor at all and it would still work with low a control weight value
>>8634020what do you use to train a lora
>>8634037ez-scripts, saar
>>8634036I wasn't thinking of using a preprocessor if there was a model that just worked. But right now actually I'm not seeing a lineart_standard. I see lineart_realistic and lineart_anime. Is there a different repo than the civitai page or something?
>>8634039>Is there a different repo than the civitai page or something?give them a try, shouldn't be much difference anyway
>getting weird double bodies in one lora
uegh i gotta sanitize the dataset more than i want
>>8634061get to work, sweaty
>>8634061Just get ChatGPT to filter your dataset.
>>8625200https://files.catbox.moe/j1o09l.png
>>8625208>>8625457https://files.catbox.moe/5ifbav.png
Love them anti-stress balls
>>8634091Box? Did you just increase the weight on excessive pubic hair or is that an artist thing?
>>8634098both
https://files.catbox.moe/ypb4go.png
file
md5: 572d817aabcf226175a49c932bc73802
🔍
file
md5: e8c10ebeb56a93c0cf7ce5aa389ff28e
🔍
can you make the picture lighter
get a new monitor instead
>>8634072https://files.catbox.moe/v2onts.png
https://files.catbox.moe/pjrebz.png
cfg++ as well
>>8625413oh boy a new vae lets try it out
>>8634141hmm, I should remade that barely distinguishable bocchi gen
>>8633377Thanks, anon. Seems most of my flat or anime mixes just don't vibe with it at all. 2.5D animu fairs a bit better. Or it just might be my brain not accustomed to not seeing a full 3D HAPA.
https://files.catbox.moe/4piegu.png
>>8634141make the tattoo glow
>>8634242I lack the photoshop skill and it doesn't come out in t2i. I was surprised the pupils even came out correctly.
>>8634192feet stuff is perfect for her
>ugly_man changed to ugly_bastard
Should we take this personally, /hgg/?
Is there a tag for when the hips/ass is visible from the top of a subject looking down their body? Like here https://gelbooru.com/index.php?page=post&s=view&id=9526725
>>8634159yeah it is very painful, just hope you arent doing anything too large scale because the trained model may not be good enough
>>8634061Ok hell I fixed it
Also, I'm thinking of rebaking that stabilizer with all the datasets I gathered now
Maybe it will work with way more images that are all more animu, I think I could get that to like 15k images lol
How do I access stuff that's up for sale on DLsite?
For some fucked-up reason, it keeps directing me to the front page. Are foreign IPs not allowed to buy shit on there?
>>8634479Yeah. Part of the creditcard co. crackdown on hentai. Certain games, think it's dependent on the content, are straight up blacklisted from countries outside of Japan. Simple proxy is all you need though.
>>8633280same, the whole thing is ridiculous. i don't understand what it's attempting to improve by doing that noodle shit
>>8634548absolutely mogged. a kritty in the making.
>>8634371Top-down bottom-up? Good taste btw.
>>8633934Male buttholes are disgusting.
>>8634481Not him but chat is this real? They actually blocked us from buying content with points? I haven't bought anything in months but I used to buy points at least and then use my coupons. I guess it's cunny only they blacklisted people from? Fuck payment processors man.
>>8634576i think you can still buy it with points if you have it in your cart its just the page itself that they blocked if you don't have a jp ip if i remember right
>>8634575I was gonna say the same thing but I knew he'd meme me and say "don't look at it, faggot."
>>8634575just give it a lick. it taste just like raisins.
>>8634574I'm actually trying to create a really (from above:999) camera angle, but that's a good idea, I'll try using that with prompt editing.
>>8634585just draw monica. do what 20loras be doing.
>>8634575>that's what he looks at in the image
>>8634371I get this when I prompt "ass" in a pov blowjob
>>8634481But that doesn't direc you to the front page, that says "can't buy this from your location" or something.
i didn't know that women had evolved dildos attached to their legs in pursuit of sexual pleasure. Truly impressive.
https://files.catbox.moe/cmexlc.webp
This whole x\y is weird overall for a concept I didn't think would need more then 2 epochs to train.
>>8634606maybe i should have let it train for longer
https://files.catbox.moe/kq7ptd.png
https://files.catbox.moe/qtmawd.png
>>8634608yeah that shit didn't work also based girl choice
>>8634613vtubers with no tits good
>>8634606>>8634608hot, I was thinking about doing something like this the other day
whats the name of the artist who draws defined tummies? its like gin or jin or something
>>8634091>https://files.catbox.moe/5ifbav.png>Love them anti-stress ballsmany thanks
>Find a good-looking lora for an artist not in illustrious dataset
>Download it
>Output fried as fuck
>Go check the page again
>All examples on WAI, all other gens on WAI
I know people training on Noob-vpred gave some people the impression that derivative models are OK to train on but holy shit at least mark your lora as being trained on WAI you retards
>>8634653It's 1206 I promise.
IT WAS 1216! ONE AFTER MAGNA CARTA!
>>8634681based futa enjoyer
>>8634681based. btw enable your adetailer bwo.
>>8634681You need to go back. Your board is so fucking full of dicks that it makes no sense why you'd need to post here. Other than to troll of course.
file
md5: f91b202f7e855d7e26fa497d9733513a
🔍
Couldn't get female pov to work but this is close enough I guess.
>>8634691Would be based if she had pussy instead of balls. It's just a tranny at this point. If you cover the top half of the image it looks like a dude, even the legs and hips.
>>8634712That's better than female pov.
>>8634712could've made that into pov deepthroat
>>8634712female pov is sexy
I like seeing tits from a first person perspective
>>8634754just forget to add 1boy to your pov groupsex gens
Aside from usual shilling, it feels so good not to inpaint hands. Perfect hands out of the box is insane, when compared to control net, manual painting, inpainting depth masks and other horrors like control net just to get serviceable hands.
>>8634576As far as I know, cunny is still being called "shoujo".
You can enable AI stuff in your searches (it switches off by itself), you can see it in search results, you can add it to your wishlist, and you can buy it off the wishlist with bitcash or whatever.
But you can't look at the product pages or look at the generator's creator page if they're filed under /aix/. Unless you get on a nip VPN, anyway.
So it's extra steps to placate the processor niggers.
>>8634782Uhh did you just arrive here from 2023? Hands have been fixed since pony.
>>8634785Yeah. Sure they were.
>>8634783>So it's extra steps to placate the processor niggers.Why would payment processors hate AI? Or do you mean people are spamming AI-generated shoujo?
>>8634786I'm sorry you weren't here back then but you don't need to be such a raging newfag about it.
what is lil bro doing to get mangled hands in current year
again on focusing inbuilt artist quality with loras
not much to say
>>8634792Who even is that? Those don't look the same.
>>8634794>Those don't look the samei mean yeah noob artist knowledge is pretty shit on closer inspection
>>8634571what did he mean by this
>>8634779sexy, cute and aesthetic
do you do a lot of post processing?
>>8634785>>8634790Are you baiting me in launching local, genning bad hands, only for you to counter me with
> Ha, look anon doesn't know secret sauce to good hands. >>86309716 digits
>>8632445bad hands
>>86328196 digits
>>86328506 digits on hand on cock
>>8633368Incoherent nails
>>8634152nails on gloves
>>86345496 digits
>>8634808man i thought i double checked my fingies
>>8634808I'm stating the historical fact that confetti and autism solved hands in January 2024. You not knowing this outs you as a newfag that should simply accept that and move on. It's really not a big deal.
>>8634808That's her shoulder, not a thumb
>>8634808Alright, I'll give you that one on my bocchi gen, I admit I didn't double check the hands on that one because I was too horny on that moment
>>8634810I am stating historical fact that I am in this thread from October/November 2022.
Pony, Autism did improve hands over SD15 mixes, of course they did.
But they are not 'solved', especially when compared NAI.
>>8634809>>8634812>>8634814I didn't mean it to attack you, I just took the bait. Sorry.
>nai goood looleee xddd
>spacing
kill yourself and go back to your containment general
>>8634818Is fine, it is a valid criticism
>>8634818You can mince words about my hyperbole, I guess, but at least now we're saying the same thing. Don't know if I'd even say NAI 'solved' hands neither but the current models have made hands a nonissue which they weren't in SD1.5.
>it feels so good not to inpaint handsLet me remind you that this is your original bait. I don't know why you'd inpaint hands on autism or NAI but this revelation you had today is not new. That was my point.
>>8634808Yeah I guess I could clean up the nails on gloves. Genned hands are a hell of a lot better now.
>>8634825I linked 7 posts with broken hands in my response. From just this thread.
Could it be that they might need to use inpaint them to fix them?
>>8634827So then let me just quote your bait post for you so you can read it again.
>>8634782 What are you even referring to? What model has perfect hands out of the box? You are the baiting shitposter here not me.
>>8634828I wasn't baiting. At least I didn't try to.
I was referring to not inpainting hands on my NAIv4.5 gens once. Or maybe like once.
Which felt good to me, when I remembered how much I suffered with hands in the past.
sooo where are the gens :333
>>8634830You're saying "NovelAI" has perfect hands out of the box? At base res too? Can you post some pics then because this is the most premium bait I've ever tasted.
>>8634830Actually let me make a request so you don't do some bullshit. Do "own hands clasped" and post it.
it gyot awfully quiet :<<
>>8634834Anon, I described my personal experience. They are not 100% perfect, but it feels good in comparison to me. That is all. But I didn't gen local since summer of 2024. Maybe new mixes are way better than autism was at the time. Or maybe my memory is not perfect.
>>8634835I know it's difficult tag, tried it a week ago. But.
https://files.catbox.moe/z190jc.png
Got it on third gen.
>>8634840Actually I see already that it is broken.
>>8634840My point was that the improvements of noob over autism is concept knowledge and aesthetics rather than hands. The same stuff we could do back on autism we still do now, just more 2d with better artists. I would have been very surprised if NAI had reached Sora's level on hands.
did you know paid services are easier to work with and more convenient than alternatives?
picrel
md5: 5625c1ee68bbac9d18191c8cc50080f6
🔍
jej
md5: 1b5efb47da47550be8ec8441eea0b73d
🔍
>>8634845damn gpt didn't know you wrote naughty words like that
>>8634846>take that cloudshit back to /aicg/bwo... I wish local textgen was decent man...
>>8634848Hey on local you might get the saas censorship but at least the outputs are shit!
>>8634805My flow usually looks like this
1. Gen decent base
2. Img2img at 0.35-0.60 denoise, 0.05-0.20 added noise, depending how much denoise and added noise breaks the image. But added noise step is usually crucial for me as it makes image less 'flat'
3. Download image and using PS Levels, Curves, Brightness/Contrast, Vibrance/Saturation adjustments until it looks better to me. For this one I also used photo filters to make it look more 'night time'
4. Use regular brush and PS spot healing brush to remove watermark if present, and remove weird items and clutter if they are there.
5. Use it as a base and upscale it with 'Enhance' in NAI at 0.35-0.51 denoise and 0.00-0.10 added noise, depending on how much denoise and added noise break the image
6. Download it and with spot healing brush remove few things if they are there and another pass with filter and color adjustments
7. Export it with lower resolution
That's about it.
man all of these image sites have really shitty quality images on them for scraping
>>8634915>resizing lorashow i do this, mister?
>>8634915then don't bake them big in the first place
come on mate this stuff barely even counts as /e/ and I feel bad when I post my masturbation pics...
>>8634933Forcefully exposing underwear is extremely erotic.
>>8634788We're locked out from looking at the product and creator pages for the entire AI subdomain. Also, DLsite stopped accepting submissions for "fully AI" products because they can't keep up with it well enough to filter out the guys trying to sell without properly tagging. I don't know what this means but "partially AI" panelled manga and CG packs are still going up for sale.
There was probably some hysterics over AI cunny last year that looked too realistic, but as far as I know, all cunny in general is still searchable under the shoujo genre category.
>>8634934But it's not hentai now is it?
>>8634933do you retards really
>>8634863interesting, thanks for sharing
>>8634922every lora baking script should have a resizer
>>8634926nyaa i bake 16 but it sometimes still helps
2/10 maybe at that size though
>>8634935Correction: Apparently most cunny is being suppressed from DLsite search. This is news to me.
I'm sorry, Anon. It's the opposite of AI where you can't even find it to buy unless you have a direct link. What a fucking clown show.
Won't continue this discussion in the next thread.
as for why to resize, picrel
it can help fix some frying in like this trying to readd her arm lol without losing style or anything
https://files.catbox.moe/q858zv.png
>>8634915then don't.
storage is cheap. anyone complaining about filesize in 2025 are retarded fagets.
and there isn't a practical reason to do so purely for file-size(layer slicing can have practical reasons but that's only reducing filesize due to how it works, it's not the reason you do it). You'll always degrade the lora for no reason other than because some retard with a tiny spinning rust drive bitches and moans about it.
he says right under a post proving him wrong lel
>>8634954Resizing really fixes that type of shit? That should sounds like random chance placebo stuff to me. It could very easily have gone the opposite way.
>>8634958>500mb glora anon as arrivedKek. You're not wrong but still. So many characters in this general.
>>8634971I mean it can, one of them just did for me but you lose nothing spending four minutes to resize shit unless you want to bake and upload to civit or whatever lol. And generally the trend has always been that it at least stays neutral if it doesn't fix it.
used this lora https://civitai.com/models/1691977/pieck-finger-attack-on-titan-wainsfw
>>8634958oh yes and with float32
>>8633691>>8633800Figured out a bit of the sauce, just needed some loras and frame interpolation.
Today I will remind them.
what is lil bud going to remind us
>>8634808>all these are local gensman, this is kinda embarrasing
naiv3 gave me consistently good hands, v4 was kinda meh and v4.5 is decent at hands but it became incredibly good at genning feet compared to v3 so it's a win-win
For comfytards
I just found this and it has made my life a million times easier than manually entering or clicking through dropdowns to get loras. Also easily accessible thumbnails for each of my hundreds of style loras makes finding the right one so much easier.
https://www.youtube.com/watch?v=hvKw31YpE-U
Figure i'd drop this here and help a homie out cause I wish I found this shit awhile back
how many vrams will i need to do a finetune of sdxl minimum
Waniggas, how the fuck do you get gens to do actually deep penetration? Shit's driving me up the wall. They just bounce on the tip or just do like an inch in if I'm lucky, or they'll just idle on top of the dick.
I've tried the General NSFW lora, the c0wg1rl lora (and other assorted positional loras), the Furry NSFW one (apparently it's supposed to be more "general purpose" than just furry), and it's always the basic bitch shit. I know these should be able to do what I ask, because there's a fuckload of user-genned shit that's working fine, and their prompts are dogshit. Honestly, at this point I don't even know if the loras are working at all.
I'm using the LoraLoadModelOnly nodes (the normal lora nodes don't seem to work), and connecting it like this
LightX2Vid (0.8 str) > General NSFW/Furry NSFW (0.8) > Sex Position Lora (0.66-1 depending on the lora) > PatchModelPatcherOrder (weight_patch_first, full load disabled)
>>8635285Bruteforcing works 50% of the time for me. I can always inpainshop if I wanted but I refuse to use a lora for that.
>>8635287You can inpaint videos?
>>8635288Was it about videos?
Isn't it just seed gacha?
>>8635289Yeah.
"Waniggas"
Wan+niggas. Honestly I guess that could have been confusing, my bad.
Images I'm pretty well versed with that shit. I hardly ever touch loras as there's not much need for me there, prompting works pretty great.
Vids though are another story. No fucking clue what I'm doing.
>>8635292I tested a little bit of wan before dropping it with a deepthroat lora. Also didn't know shit about what I was doing.
50% of the time the girl would do nothing but a have seizure on top of the dick.
40% she would take the tip or something
Around 10% was the actual deepthroat success.
I always assumed people cherry pick their best video gen and that's what you see.
>>8635296>I always assumed people cherry pick their best video gen and that's what you see.Seems kinda shitty considering how long it takes to gen a vid. I'm averaging about 200s, but still just seems like too long of a process to rely entirely on gacha. Seems like something that would have people really getting their shit together to gen things right the first time.
Still, good to know that it's not just me having problems getting shit to work. I'll admit, it does work 'sometimes', but I figured there was something up on my end that I could do to nudge it in the right direction.
in reforge when you use the weight hotkeys on a two word tag it does this (huge:1.1) breasts
any way to fix this?
>>8635302just select both tags before using your hotkeys bwo
>>8635302I've never had that issue.
Stupid question, you do have both words selected right?
>>8635303>>8635304yes i know you just select both words but its not optimal and bad logic
should be that it just weights up everything between commas
>>8635306>should be that it just weights up everything between commasI'd rather just select everything I want (which may contain multiple tags and commas) and weight them all, than weight each tag individually. Hopefully you're using the general text editing hotkeys of ctrl+shift+left/right arrow to select entire words at once (shift+arrow for single characters), makes that shit super fast. You can delete whole words at a time with ctrl+del/backspace
>>8635308i never said i wanted the selection feature removed albeit
You know, given that all of these style loras also fix the high bright-dark contrast in pictures, I wonder if it's stabilizing the vpred idiocy, stabilizing the eps merged into 102d, or something else altogether.
test
md5: 0df5a9a39464bbbefcbbd3aa406ed702
🔍
>>8635333The former. It also removes most of the advantages of v-pred, at least when it comes to lighting. Doesn't happen when prompting artists, only with loras.
>>8635340Can't say I mind, contrast is a meme desu.
>>8635274finetune or lora? 16/6gb?
>>8635285>>8635296Well I think I have the issue solved... I think.
The FusionX model mentioned in the Rentry is working wonders. Both for extra movement and for actual sex.
>>8635350I don't mean contrast but being able to have a dark scene in front of a dark background. Or not getting local lamp in every indoors scene.
>>8635340Zankuro hag hits different
>>8635380Eh that's what I meant too, sorry. Honestly though, dark hentai scenes don't even register in my mind.
Today I want to train. Is training a noob vpred lora at 1536 res a good idea?
>>8635408Nah, a higher res will only introduce errors with literal zero actual quality difference. Trust me I tried.
>>8634574Btw it did work, but it doesn't quite get the camera angle I want so I looked for other solutions and found a lora surprisingly
https://civitai.com/models/1424057/il-pose-from-above-head-top-view
otherwise I probably would've tried controlnet depth from blender or something.
>>8635302I've always just typed the stuff manually. Habit.
Any way to do tongue_out but sticking straight out towards towards the viewer rather than lolled out?
Like this. https://danbooru.donmai.us/posts/4871111
rather than https://danbooru.donmai.us/posts/4884665
Might try controlnet but still.
>>8635402Will you post some results too?
>>8635408better to just add some close-ups
helps when upscaling
is there a reason stable diffusion stops working after 300 in the prompt? i might be retarded. i never had this problem but i recently reinstalled windows and it's been doing it ever since
>>8635243>naiv3 gave me consistently good handsYou and I both know this was artist dependent. What are you even trying to prove here?
>>8635340>only with lorasGlad someone finally came out and said it. Thought I was the only one having this issue.
>>8635387Ah, nonnie is a 'lights on' type of guy huh? Still I'd love to do a cowgirl scene at a matsuri with fireworks going off in the starry sky and not have it look like it's noon.
>>8635415>>8635408Wasn't some anon saying it was better at like 1200x1200 or something?
>>8635419Weird. I know for a fact I saw anon post a pic recently of a blowjob with the girl on all fours and it was directly from above so it should be possible without CN or loras. I guess you want a solo though? I'll try to make one when I fap next.
>>8635528I think it's just conditioning kek, I don't remember the last time I saw a real night hentai pic.
>>8635533I tried baking at 1256x1256 (for use in that no less) and like I said, zero additional detail and it just introduced burn in errors.
>>8635554Nice pic. Simple but effective. I really have to stop doing pov.
file
md5: 73559144f44a2cf27d582f61eee481c2
🔍
so how's chroma these days? Can it be at least used to generate a dataset for a style lora or something?
>>8635556That's cuckold pov...
>>8635557Why would you use chroma for the dataset?
>>8635558if that's what you like, who am I to judge?
>>8635558nice, netorase is the thinking man's fetish
>>8635588>thinking manwhat exactly is that man thinking of`?
>>8635590thinking of (You)
Why do grey loss decreases all the time, while the rest stop after a while?
>>8635636the fuck is the grey loss
>>8635649>>8635644Grey is just one picture dataset and peaked at ~300 steps. Blue is 2+6 rotated/cropped pics and 16/16 dim/alpha, pink is the same dataset, but 64/64. Bs is 1.
Because of 400 steps, I expected to lower lr for blue and the best result at 6000, but the output lora had almost no impact.
Pink is the good at ~2000, but a little bit better at 6000.
>>8635656>one picture datasetwhy...
Should I Take A Chance on This?
https://arxiv.org/pdf/2505.14238v1
Styles
aaaa \(quad-a\) https://files.catbox.moe/txc9o2.safetensors
bacun https://files.catbox.moe/nl7j2m.safetensors
celebelian https://files.catbox.moe/4a29ww.safetensors
chalseu https://files.catbox.moe/rmcmgz.safetensors
combos & doodles https://files.catbox.moe/r6b10l.safetensors
dyun https://files.catbox.moe/gs68zj.safetensors
ikeda jun \(mizutamari\) https://files.catbox.moe/7rlmtz.safetensors
ikujitto https://files.catbox.moe/woaimd.safetensors
inoue takuya https://files.catbox.moe/i1jrzp.safetensors
jishaku \(user rcuz2843\) https://files.catbox.moe/f2ky76.safetensors
kanesada keishi https://files.catbox.moe/ro5vmm.safetensors
kotteri https://files.catbox.moe/muy5k7.safetensors
marimo https://files.catbox.moe/mmh7wo.safetensors
midou tsukasa https://files.catbox.moe/3llmfk.safetensors
murasaki iro https://files.catbox.moe/95f1mv.safetensors
nanamo \(002a\) https://files.catbox.moe/rq33n8.safetensors
narisokonai https://files.catbox.moe/iqznas.safetensors
nnn-ollll https://files.catbox.moe/decifp.safetensors
nona moth https://files.catbox.moe/qsaae5.safetensors
orange brand https://files.catbox.moe/y92jw2.safetensors
riffraffhellyea https://files.catbox.moe/x2r6ev.safetensors
riz https://files.catbox.moe/pl6m4e.safetensors
sape \(saperon black\) https://files.catbox.moe/my2160.safetensors
sawa jaaji https://files.catbox.moe/hqkpft.safetensors
suoiresnu https://files.catbox.moe/9iny9u.safetensors
tsuda akira https://files.catbox.moe/w4a3p3.safetensors
yan \(bltyann\) https://files.catbox.moe/fcpzxr.safetensors
yn red https://files.catbox.moe/mh5e02.safetensors
>>8635683Wait, it's actually three pictures dataset.
I still don't know how to bake a lora for a character, so I don't have to enter all the colors for hair and apparel.
>>8635704None of these are painterly though.
>>8635704I didn't recognize any of those so I took a look
Wish I didn't
>>8635741quad-a is okay, not sure what's the point of the lora tho since he works fine by default. Used to be a decent anime stabilizer back on Pony too.
>>8635757Used to be one of the very very few styles that produces decent anatomy on Resonance Cascade...
>>8635757>Used to be a decent anime stabilizer back on Pony too.Really? never heard about the artist before or being used that way
>>8635757>not sure what's the point of the lora thoeh these models don't really do normal lighting, see all the examples above
test
md5: 24b79826282252df87435a9ff80d2f13
🔍
>>8635763I think because most people used loras instead. Better than asurauser anyway, at least you don't get the /aco/ bias.
I was thinking, for the "stabilizer" lora, would it be better to keep the artists tagged or not? I assume not.
>>8635849So the idea is to just have a very weak lora that gives you a slight bias towards flatter styles than "masterpiece, very awa" would? If so better not, so that it doesn't require a trigger tag.
>>8635857Yeah I want to bake something to stabilize noob
Previous one sucked but now I have way more quality styles haha
>>8635864seems pretty stable already if you just prompt an artist and keep the CFG low, around 3
I don't even use the CFG++ snakeoils anymore
>>8635866maybe if you really don't like how the artists look desu
most of them barely resemble each other
>>8635704do this style, thanks
https://www.pixiv.net/en/users/141568
>>8635849For your stabilizer can you also make it drive away from banding and splotchy black levels?
>>8635926We don't have the tags for him, unfortunately.
>>8635943I dunno if it'll even work but the one I baked on a shittier dataset somewhat did. This dataset is 1k images and 90% anime so it should? IDK, I'll see.
fuck you guys were right about the lamp wtf
>>8635702anyone got a config for this? usual lycoris one just produces a garbled mess
Why do people end up putting up sets up for sale that are all out of order?
Like, the images are jumbled to shit but if you (the buyer) spend time sorting through by state of undress and sex position, you can create a scene with continuity to it that's considerably better to flip through.
In any case, I'm thinking of doing this with stuff I like enough to put on the Panda but there's no tag for it that exists. Good idea or bad?
>you can (not) prompt for different directions for tongue_out to point
Damn, maybe it's blendering time.
>>8636179just draw a red blob, why would you need blender for this
>>8636161What do you gain out of posting AI on panda of all places? Shouldn't it be an archive for pirated works?
>>8636187Because I want to gen tons of different images but with the same pose. Maybe if there was an automated workflow that places the red blob with AI.
Anyone up for nitpicky highly specific request involving Pregnant Pokegirls?
My concept is cool so you might find the process fun.
>>8636212pregnant cynthia beach scene, not sure how nitpicky you want exactly, but just do whatever
>>8636212pregnant Selene but she is too retarded to understand that she is pregnant
>>8636223>>8636227No I mean -I- want to make the request lol
>>8636240Im fucking idiot, fire away then
>>8636240Then just post it, at worst people will tell you you are a weirdo no need for all this foreplay
Requesting 18 sexy/cool pregnant pokegirls 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
>>8636257k nvm, I aint doin that
>>8636263Kek I imagined not, it's quite autistic. Welp fingers crossed.
>>8636271the /slop/ thread on /trash/ may be more accepting of your request
>>8636273Thanks. Are /b/'s degen threads good too?
>>8636257>no Hilda>no Rosa>no BiancaShit taste and easy skip.
>>8636275meh id skip them
>>8636276-I- didn't decide on the pokgirls he gen'd. All I asked specifically for in the way of pokegirls was Electric Lana, because of the 0k0j0 doujins
>>8636283Eh I'll give it a go anway. Loli is the one thing good about /b/
>>8636188I'm still retarded at making my own and want to promote good generators by pirating shit I like a lot.
Well, "good" as in they know how to create content even if they don't understand how difficult it is to masturbate to a 100% discontinuous series of sex scenes.
>>8636257I really don't understand the kind of mentality someone can have that they'd think they'd get any of this shit for free
go pay a commission fee. no one is doing this shit.
>>8636341Well one guy did it and it looked pretty decent, and I figured 1 small step up of effort from that trying to actually match my /vp/ posts was pretty likely, though obvious lessly so in precise proportion to the step up in complexity I'm intending.
I think I'm feeling a little more optimistic about my chances than you.
>commisionThat's step 2 if asking you pros doesn't produce a high enough quaity result.
My angle is that I'm hoping to fish in someone who thinks my idea is as cool as I do, and does my request in appreciation. I'll have to wait, doubtless.
whoops
>obvious lessly
*obviously less
>>8636341don't see what the harm is in asking first
ok nah stabilizing base is still a meme
once again i am staying with 102d
>>8636364your "stabilizer" is literally just a lora that isn't shit trained on an artist style you want to be using.
there's no random "generalized" nonsense to it. It's just a fucking style lora that isn't trash garbage.
this isn't complicated.
>>8636376but is he also wrong
>>8636161You mean the people sharing like 3000 images per week, with zero care given to curate them? Basically just dumping their "out" folder.
I don't think they give a shit about quality.
Little did people know there's a sekrit noob vpred version continued from eps 0.5
>>8636257Fire pussy in a cabin in a blizzard sounds peak fucking comfy.
>prompt downward dog from behind
>it's always low angle shots
>inserting from above at any weight makes it not do the downward dog anymore
>doing [from behind:from above:0.2] almost works but doesn't
Ahhhhhhh blender/controlnet tuskete
>>8636479facing away, sweaty
>>8636479>downward dogI had always used frontbend for this
>>8636482How does that help? (it doesn't btw)
>>8636486I pictured AI being unable to do from above because you prompted her eye color and such since most of the pics with that tag is view between legs.
>>8636490>you prompted her eye color and suchI did?
freaky ahh schizo moment :skull: :skull:
>>8636485Are you using some other tags to reinforce it? I tried that tag in isolation as a test before and the model (base vpred) didn't seem to know what it is. In the end a combination of "bent over, downward dog, frontbend, all fours," worked and is what I'm using at the moment. I also tried bear position but it didn't really seem to help and actually it sometimes gave the girl bear kemonomimi kek.
Why lil blud prompt her eye color fo?
what a hideous of writing
where the fuck is 4.5 inpainting/vibe
>>8636595>4.5 inpaintingdone
>vibewaiting for it too
>>8636596oh cool. have you tried it? how does it compare to v3? I was waiting for that to resub but might wait for vibe too if there's any estimate on release.
>>8636419https://aibooru.online/posts/123316
>get quite a handful of really good gens by this point that cater to some niches and has a unique art style too
>suddenly feel the urge to use them to attention whore and maybe patreonmaxx
>moral values prevent me from acting on it
>>8636747i'm sure whoever asked is really proud of you.
>>8636749uh.. puffy_anus bros? Our response?
I've regressed to genning pantyshots
>>8636814we lost another one to the /e/ menace
>>8636814>be on /e/>want to gen hentai>be on /h/>want to gen ecchi
>>8636833>be on both>and /d/, /aco/>want to gen nothingbut I still open the UI every day
I saw a few ST threads on /v/ recently and got baited into trying Deepseek R1. I hadn't botted for a good 6 months as I grew tired of local 70/72B slop.
Bros.. I haven't genned since. My cards are doing shit I never seen them do.
>>8636853You're running a 670B model?
>>8636853>Deepseek R1Did they fix it being a total schizophrenic with literally any preset?
>>8636864Nah. I sold my call center so I'm using some creds from OR
>>8636865I'm unfamiliar with Deepseek as a whole prior to a few days ago. Well, I knew of it but never used it. Been a local slopper for the longest. But what I tried myself after some rentry preset hunting? Great. I'm sure it'll wear off as it always does but for now, I'm running a ton of my custom cards through it and they're just surprising me.
>>8636671Not yet, I'm waiting for vibe to resub for a month again.
>baked a lora fully on scraped doujin without cleaning or anything
>it actually just werks
I wouldn't say it looks just right but that I get zero panel leaking or anything is neat.
>>8636883well actually prompting it in actual grayscale it's pretty accurate
hm maybe i should go on a little doujin bake run now
>>8636884do yourself a favor and do a minimal clean up pass at least
just because it work on your few test doesnt mean it wont break apart with further testing
Speaking of text coom, sao10k-32b-qwen2.5-kunou surprised me positively just now. Really good dialogue and smut capacities, doesn't go on weird tangents or anything. Keeps the plot proper.
Kinda retarded knowledge wise but that's to be expected I guess.
There's so many shit artists and shit art out there, the limitations of current models don't feel that bad desu.
>>8636934That thing's old. Have you tried the new Mistral Small 3.2 (use tekken v3 tokenizer settings according to /lmg/)? There's also Gemma 3 27B which has the most knowledge at that size but might have some other issues.
>>8636937I never liked Mistral back when I was in /lmg/. Gemma is okay but it's definitely just pure GPTslop instruct. Useful for tasks but the coom is meh.
>>8636935I guess but the composition is a real issue
>>8636739Nice. I don't like this particular style of painterly (looks like digital brush instead of traditional), but composition and detail are very sugoi.
>>8636853Yeah what I do these days is go on an adventure with my waifu and then gen pictures of whatever's happening in the scene.
>>8636865Yes.
>>8636973>Yeah what I do these days is go on an adventure with my waifu and then gen pictures of whatever's happening in the scene.Based effort anon. I remember doing something like that when I used NAI a few years ago through some extension on ST. Now a days I feel I'm too jaded from my time in hdg and hgg that the first gen I'd need to toss into photoshop to fix minor errors, then inpaint for a few minutes, then upscale for a few more. It was so much simpler when I accepted any slop.
>>8636976Well that's the thing, I always remind myself that I'm gooning and so the first pic that "looks right" should be used. It gives me a chance to learn danbooru tags/practice since I have to do it as fast as possible to not lose my boner. The only problem now is I do too much pov.
>>8636978>The only problem now is I do too much pov.Understandable. I've actually settled on going the full adventure route and then genning the final scene. That way I can give myself a bit of re-rolling autism leeway if need be.
I kind of want to give audio a try. Again, only time I tried it was a few years ago through NAI's audio bots and it was awful.
>>8636978>The only problem now is I do too much pov.Taking the disembodied penis pill is the best thing you can do for yourself haha
>>8636996Yeah you're not wrong kek. I always forget in the heat of the moment since "penis" is a lot faster to type than disembodied penis.
Is generating samples during the training making any sense? For all of my bakes, they all looked completely fucked, but when I use the resulting lora on the same training base model checkpoint, it looks fine...
You know, I think baking artist loras also helps with lowres melties. I'm testing all those bakes I did and I think they look way better and cleaner at lowres (well, 832x1216) than their respective native versions. I think I'll finally switch back.
>>8637026I don't think there's much point when you're just going to run tests on the end epochs.
>>8635704man on a mission
I should do the same with the artists I wanna fix up
>>8636814ai has been feeding my pantsu fetish for 2 years already and it's still not enough
>>8637039once you have a good config it's pretty easy
just waiting
>>8637044gathering datasets and testing epochs is still annoying but yeah you're right
>>8637046best thing i did for this testing was enable save only last x epochs, cuts out so much bullshit
it was worth testing when this setup converged
as for datasets honestly just use a scraper for danbooru, fuck cleaning up r34 garbage, the one i got from gpt can scrape like 300 pics in a minute
>>8636814>>8636826I don't usually fap to my own gens but when I do it's mostly to an /e/ gen
bros, the artist i want to train on draws shit hands. wat do? is masked training or something like that used in situations like this?
>>8637056crop out the hands or inpaint them
>>8637060>inpaint themi guess i'll try that. dataset preparation never ends, huh. thanks anon
>>8637051Sex is the lowest form of eroticism t b h.
Uegh I wish hiresfix wasn't so sluggish
>>8637117latent is faster than using esrgan upscalers
>>8637155>he's not using x1.5 latent + x2 tiled CN upscale
>>8637155maybe but genning base higher res is like twice as fast
>>8637162 me
oh wait i just realized highresfix doesn't scale the upscale steps like denoising does for whatever reason
this will be faster now
Has anyone compared normal hires with the kohya one?
>>8637177the deepshrink? i can test that i just downloaded it
>>8637191Yeah. I don't know how to get the settings right for that or what the theory behind that stuff is. Like I adjust one setting and it fixes some issues but then at other aspect ratios I get other issues.
>>8637219nta but just change the end percent to 0.4 instead of 0.35 and leave the rest at default
file
md5: 9070e7e7235cea82a170d846ee023ea8
🔍
>>8637224yeah, I didn't test it that much but with anon's high res lora it was possible to gen at 2048x2048 with those settings, or 1536x1536 without it
>>8637228what is kohya hiresfix supposed to do?
>>8637233won't claim to understand exactly how it works but afaik it just gens a lower resolution until the end step at which point it starts genning at the set resolution instead and the result is a higher res image without the problems of just raw genning at that resolution
>>8636419https://files.catbox.moe/yl5x4r.png
>>8637228>>8637236Hmm alright, I'll do some tests on high base resolution
Also regarding hires upscalers
I don't think you actually need one? https://files.catbox.moe/9wb0kh.zip
these look identical detail wise to me, 0.5 denoise at 1.5 same seed
>>8637330They don't really do much of difference once you add i2i on top.
>>8637333yeah literally none other than fucking up colors
huh neat, another save
i think i thought the opposite before because my sampler/scheduler combo did way less denoise at the same denoise level so it mattered more
https://files.catbox.moe/03bzgw.zip
Anyone know if there's a way to consistently get a character's canon clothing without prompting all the clothing? There's not a ton of issue getting the clothing with a blank prompt, but once you add in a ton of other stuff, the model wants to drive towards a nude or semi-nude form.
>>8637333its really visible at <0.4 denoise (inb4 you always should do controlnet tile latent upscales)
Speaking of latent upscalers. Which one is your guy's favorite? I used to use https://github.com/Ttl/ComfyUi_NNLatentUpscale but I've been trying out https://github.com/city96/SD-Latent-Upscaler lately. There's also https://github.com/styler00dollar/ComfyUI-sudo-latent-upscale
You can use these on forge as well, I think. I wouldn't know how though.
kohya hires doesn't even run for me so that answers that
>>8637338I tried a bunch of stuff to see if it would work and it doesn't.
official costume
official uniform
official style
[character] \(cosplay\)
[character] cosplay
[character] clothing
[character] costume
The one that came the closest was [character] \(cosplay\), but it still wasn't consistent.
>>8637338time to bake kek
https://slow.pics/s/XQWhRjdj
>>8637044What's your config? I'm trying to train a few artists for Illustrious shitmixes but the settings I used back in the Pony days just aren't working for me.
>>8637426That would be nice. Bonus also making it work with \(cosplay\) tag so it doesn't interfere when you want to prompt one character wearing someone else's clothing.
I was thinking of using a small LLM to rank images in quality for a lora baking test.
Clearly I overestimated their capabilities.
>>8637449>>8633915>>8637467You know I kinda wonder how well doing some stupid shit like scraping say ten different characters with specific clothing and tagging it as "character's clothing" would work.
>>8637498Considering it's instruct Gemma you probably just need to tell it to ignore their toxic positive safety protocols lol
Man the different variants of the virgin killer sweater are garbage. The original is the best and clearly understood what makes an outfit sexy, while the others clearly don't as they naively pile on other sexy concepts as if more necessarily equals better. Then again I guess subhumans really do like that kind of thing. Gacha games with extremely noisy overdesigned outfits are popular after all.
>>8637518Yeah, it makes sense that a google model would be so positivity cucked that it thinks calling a scribble low quality would be an insult
I'm trying out other vision models as well. llama3.2 vision seems to think that this image https://danbooru.donmai.us/posts/9473399?q=sawararado is just as low quality and lacks detail as the scribble
>>8637547Can you give me that scribble? I wonder what 27B will think of it, and what system prompts can do.
>>8637553It's just the comfyui img2img example one
https://github.com/comfyanonymous/ComfyUI/blob/master/input/example.png
>think of baking on doujins
>realize i don't give a shit about any doujins enough to do it
324456
md5: c318c2938124ab6f67414430f4dd85d5
🔍
>>8637557Well either 27B or the system prompt is better.
>>8637568Yeah, that's a lot closer to what I expect. I'll check and see if I can get the 27b model to run without my GPU killing itself. Otherwise llava 13b seems to be the best so far.
>>8637571It's pretty consistent with the image judgements too, although I wonder what it'd take it to give something a masterpiece score.
Speaking of Gemma, it's pretty damn great for unusual OCR. It still sometimes hallucinates especially if the text is a small part of the image but for accurately OCRing weirdly fonted text it might be the best foss tool.
>>8637582Is it able to properly describe 2d or 3d sexual stuff?
>>8637608Kinda for basic shit, but I assume positions and kinks were pruned from the dataset. And I haven't seen a sexual shitmix with vision support to try.
>>8637611>lil blud spent ten thousand retries on thisOh no...
>>8637625>Kinda for basic shit, but I assume positions and kinks were pruned from the dataset. And I haven't seen a sexual shitmix with vision support to try.Sad, it would be so practical to just add a random hentai image and have it describe everything.
Even for indexation.
>>8637639I wouldn't be surprised if adding some examples would magically fix it. I should try asking it for tags.
>>8637026>>8637034I always used it as a canary back when I used prodigy as my optimizer. If unet blew up, I'd know 30% into the bake instead of at the very end.
>>8637547>>8637498Just use free gemini.
>>8636257I like pokegirls but you sound insufferable to work with. If you are so picky, just fucking generate it yourself.
>>8637744that is one zesty ass post my nigga
Speaking of zest, where's the trap (formerly otoko_no_ko) anons at?
>>8637757Nah. Been here since #001, nigga. We have trap (formerly otoko_no_ko) anons.
>>8637750not that I posted that many, or even gens in general, but I've been here since the start
>>8637760>>8637764>I've been here since 2003pwoof?
>>8637750Hopefully dead or soon to be dead
I have several things I would want a lora for that I have encountered in my short time returning to image gen, in case anyone wants something to try, as I've decided not to get into training myself...
>horizontal/vertical/round inverted nipple control
>tongue_out curvature control (curves down, curves up, curve to side, no curve/straight/erect)
>specific camera angle control using numbers (impossible to do with the current TE?), maybe would have to use some SLAM algorithms on images, or maybe take some screenshots from a game engine that lets you get the camera's angular values
>side-tie bikini top (vpred gens front-tie instead)
>suspenders hanging (vpred confuses this for normally worn suspenders most of the time)
>crop top/shirt overhang where the cloth hangs loosely (vpred a lot of the time makes the cloth hug the breasts a bit as if they have some elasticity)
see
https://danbooru.donmai.us/posts/9228072
vs
https://danbooru.donmai.us/posts/9110972
this can be fought against a bit with prompting but not reliably.
>totally smooth, second skin-like impossible skintight bodysuit/clothing that is reliable no matter the art style (vpred can achieve this when prompting certain art styles, but on most art styles, it gives the clothing wrinkles; prompt editing using the bodypaint concept with paint/sweat negs kind of works but not reliably)
>>8637776Are you the downward dog anon too?
>>8637780Yeah. I'm also regular short person relative to environment anon too. Another reason I wanted to get into Blender. For now I'm delaying that as I feel more compelled to keep exploring base model capabilities and limitations.
>>8637711I know, hence the "tolerate". I'm not demanding delivery, I'll be lucky to get it.
>fucking generate it yourselfSimply put: I want the result of the hobby without investing the time needed to wrangle a local model well. I get you might view this a impertinent, and you're probably right, but I'm banking on coming across an anon who likes the idea enough to want the gens himself.
After 20+ threads of no delivery I probably will have to just pony up the money, just like someother anon said.
I'm just hoping I won't have to. Fingers crossed.
>>8637808At least you're self aware and not a jeet faggot which I previously thought you might be. Good luck on your /r/
>>8637750>>8637760>>8637764If you mean the Inazuma & Meshi gens, then yes. Been distracted with chatbots lately to gen thoughbeit.
https://files.catbox.moe/pgglxl.png
Was there ever a vpred CN for noob? I'm sick of the eps version washing out the colors.
>>8637668So... generated samples aren't supposed to look warped and melted as fuck? Because they totally are for me, regardless of the epoch.
But when I use the resulting Lora - it works just fine.
>>8638041>So... generated samples aren't supposed to look warped and melted as fuck?Generally speaking, no. Are you sure you're using the right model for the associated LORA? If it looks really warped you might be using the wrong base model for your sample outputs. Either that or not having quality tags.
>>8638046>Are you sure you're using the right model for the associated LORA?Huh? I can use a wrong model for that? Easy scripts only let you pick the sampler, sampling frequency, and the prompt file.
>>8638050Hm, I use Kohya GUI scripts so I couldn't tell you. It sounds like the program just takes a few parameters and runs your prompt through the given model though. Maybe it's a precision conversion problem? Like if you're training in FP8 to save VRAM but saving in BF/FP16 then it could be that your training LORA is at a much lower precision and thus has way more loss than the saved LORA, but again I'd imagine if it's a per-epoch thing you'd save the epoch in full precision and then run the sample generation script. What about size? If the scripts don't let you specify the output size you could be generating at a size that's too small for your model, that'd be my only other guess.
>>8637808>without investing the time needed to wrangle a local model wellNovelAI is just $25 :3
lil bro thinking anybody will gather datasets for him
yea but how bout lil blud?
>>8638061From my understanding, it should use the same base model and the lora network that it uses during the training.
Also, I train in bf16 mixed, no fp8 base bullshit or full_bf16.
The size is okay, 832x1216.
Before, I had problems with training - basically none of the public configs worked, either coming out underbaked or getting fried in literally 2 epochs. So there might be something with my setup...
Hey guys, could anyone check
>>8637879 ?
tldr can anyone post some reliable toml files (in their experience) for training on noob? both eps and vpred if possible.
>>8637449Could you post your pony toml? I wanted to try and bake something in it nowdays to mix with another style, but I hopped on imagegen too late and wasn't there when Pony was current.
file
md5: fa93724dce3f964676621f70282309b7
🔍
are there any loras for adding faceless crowds to your gen? preferably faceless
prompting solo focus and the various tags like; people, multiple others, crowd and what not works
but I feel like a lora to nudge it would make it better
anyone have recommendations for character loras
>>8638417What do you mean, like which are good.?
quick doujin scrape bake test, shiba nanasei
stuff really is easy nowadays, i remember people in 1.5 days and such would clean everything and still get garbage, this wasn't cleaned at all and there's no panel leak or anything
https://files.catbox.moe/o88vep.safetensors
>>8638419I think he's asking for lora baking help, probably? Or he's tired of the millions of characters on Danbooru and Civitai, but that seems unlikely. Still the question is far too generic.
>>8638419didnt realize it was unclear my bad LOL
was wondering if anyone had recommendations as for things to do/not to do when baking a character lora opposed to baking a style lora.
>>8638441Not much. Detailed clothing needs higher dim sizel and you can optionally not tag some of her features. That makes them non-removable parts of the character that show up as soon as you prompt her name.
>>8638445>Detailed clothing needs higher dim sizeWhat do you mean by this? And by that I mean which sizes and does high dim allow for difficult concepts like tattoos?
>>8638449You'll get some tattoo either way, but the shape will keep shifting with every gen if you don't have enough size.
>>8638454>but the shape will keep shifting with every gen if you don't have enough sizeWhat size do you recommend? Do you have a picture of "detailed clothing" so we can compare?
>>8638456I needed more than 40 here https://civitai.com/images/73915704
Plain schoolgirl clothes would be fine with 8. Most will be somewhere in the middle.
>>8635704Please do hamada yoshikazu style for noobai v-pred or whatever. There are decensor edit on sadpanda I can get one if you need
https://www.pixiv.net/en/users/2675549
>>8638563make the dataset urself if you want, i can tag it
>>8638587well and bake it ofc
>>8638587>>8638589since all source images can only be found on pixiv and they're in jpg format and most of pic are about 1000px to 1600px, is it good enough to do dataset?
>>8638681nta but 1 megapixel is ideal
jpeg depends on compression level, if it's too bad you might get artifacts
>>8638563Damn, I haven't kept up with this guy since he started the isekai arc or whatever, he really went hard into the coomerism.
file
md5: a24c20a3454445f3d5109f0221d0503d
🔍
>>8638716Oh nice. That will be my wife's sci-fi outfit if I ever get around to that story.
is there a way to prompt consistent touching penis with one finger or do i need to bake
>>8638773>danbooru doesn't even have a tag like thatfucking useless site
at least r34 has it
>>8638773"poking penis" probably.
>>8638785>third way to prompt for a body part touching anothersigh... but i do still need to bake it apparently
>>8638787yeah, no way a concept with 100 pics is gonna work reliably
speaking of traps i think danbooru is unbanning some artists
if you go to ikemeru19's page you can see that it's still marked as banned but the newest posts are unbanned
https://danbooru.donmai.us/posts/9470000
there are other artists too like this one https://danbooru.donmai.us/artists/103012
i will choose to believe that the admins have seen sense and are finally getting rid of the artist ban system instead of what probably actually happened which is just some images slipping through the cracks
>>8638783ahem they do
it just has less then 100 posts
>>8638773finger on penis is an e621 tag but it really didn't work that well when I tested it
>>8638789what lmao
i baked ones with 30
>>8638789why does it even need to be """reliable?"""
are you some kind of wildcarding faget or something?
English 101 is required for posting here, bros.
>>8638799NTA but it's rewarding to achieve a difficult concept with pure prompting. Some people treat image gen as a kind of game that they play, not just as a tool to get some final result.
>>8638814consistency still wouldn't "matter" then. Get a proper result once and bang zoom you're to the moon.
the only people that care about a tag working """consistently""" are wildcard sloppers.
>>8633394Okay I used this LORA for a bit but it sucked so I made my own. Hatomame Ago trained on Illustrious 1.0, so it should be compatible with non-VPred models (Sorry I use WAI instead of Noob): https://mega.nz/folder/SBJGgajC#MPVeqp_xFCjnjeaMqqTydQ
Feel free to test it and give me feedback, if it sucks and someone else wants to give it a go I can upload my dataset but I'm reasonably happy with the outcome.
>>8638223Sure. I used Kohya GUI but here's a config file you should be able to port over to your training scripts of choice. Last results on Pony were a tad bit overbaked so feel free to either dial down the LR or just use an earlier epoch. It worked fine on it's own but had small difficulties with other character LORAs unless I knocked it down to 0.8 strength.
https://files.catbox.moe/ct9mg7.json
>>8638820For me a prompt/concept isn't "solved" until the prompt hits at least 50% of the time or so.
>>8638820>consistency still wouldn't "matter" thenBro I just said the fun is in the challenge so obviously consistency matters there because that's part of what makes it challenging and rewarding knowing that you've made the best prompt that is possible.
>wildcard sloppers getting defensive when they get called out
lmao
I don't even know how to use wildcards btw
is this what the kids call a strawman
>>8638587>>8638696Ok, now I got an images for dataset. Do you guys think I need to remove or add some before I try to edit text/artifacts/censor?
https://gofile.io/d/VZf2S8
>>8638896You don't need to edit stuff out, see the manga lora above. Zero cleaning, just werks.
Although definitely dedupe stuff and select like 150-300 of the best ones. (DESU, I have some doubts on how color+monochrome training will work, those doujin loras did need to be prompted greyscale to look *fully* like them.)
>>8638796I meant danbooru pics count, not for lora dataset
>>8638911ah yeah for sure
>>8638830>trained on Illustrious 1.0, so it should be compatible with non-VPred modelsNon-noob models you mean. WAI-nsfw also merged in noobai, even if he doesn't declare it anywhere. You can tell it recognizes e621 artists that even illustrious 2.0 doesn't.
>it's penis poke on gelbooru
ooh madonne
>>8638921>WAI-nsfw also merged in noobai, even if he doesn't declare it anywhere. You can tell it recognizes e621 artists that even illustrious 2.0 doesn't.Interesting. Good to know, thanks. So does that mean WAI-nsfw is a V-pred model and I should train accordingly?
>>8638929It's more that vpred loras work very well on wai. Wai is just a strange model.
>>8638921>WAI-nsfw is also merged in noobaialright need to check my eyes
it's over
there's just too much interference with other tags i think
if im using wai v14 should i train loras on top of it or just base illust v1
>>8638947Try both if you have the time. I would expect one trained on the base model to work better, but yeah WAI is a bit weird.
And if you also want to share the lora, it should be based on illustrious 0.1 or noob v-pred 1.0.
>>8638947>if im using wai v14 don't
>>8638954i get more coherent results with it than i do noob lol
>>8638957thats how overfit slopmixes work
>>8638958greek letter model suggests otherwise
>>8638957anyway who the fuck knows anyway when the shitmixes are full of both vpred and eps at this point
do we even know what the base for wai is
>>8638790>speaking of trapsFormerly otoko_no_ko
how does saucenow work? it doesn't trace my pixiv posts. I ask because I see people repost my stuff and curious if that could lead to people looking up the source.
>>8638994It doesn't work on AI slop, sweaty.
I tried baking a Magion02 lora with
>>8633915 this toml but I ran into the same issue as with my original config, this particular style doesn't transfer. Especially not the sketch-like shading.
I saw another anon higher up talk about detailed clothing and tattoos needing higher dim size. Now, I'm a total noob at this, but it sounds like higher Dim size = more consistent detail. Could that help with the shading?
Pic for comparison: https://files.catbox.moe/4usl1d.jpg
I just can't seem to get it right, bros.
>>8639023The problem with dim is that you'll have to adjust alpha and if you do you'll have to adjust learning rate. The whole thing is just a pain so I abandoned it.
>konan puts out 100's of quality loras consistently
>anon struggles with 1
Telling.
>>8639031with that level of sorcery he should call himself Thulsa
>>8639031>konan>qualitylol
lmao
also not sure why you're trying to shill yourself here.
wasn't it "conan" and wasn't it a detective?
what is an alternative to CIV? they are deleting anything loli
file
md5: 41170286860ee44067c7e25ce6d82f1c
🔍
>>8639023IIRC the network's Dimension is basically how many neurons you're adding to the model, and more neurons can store more information. This is why a higher Dim is important for really detailed clothing/character design, because fewer neurons means the network won't have inputs for all the little details. Remember, training a LORA is basically asking the model to create an image from your training data, then scoring it based on how close it gets. Shading and subtle aspects of an art style can qualify, but it's important to remember that you're still working with a base model and all of the biases and habits of that model will still show through. Like if you train on an artist that is known for wide shots or unusual body shapes, a model like Noob/WAI is still biased towards 1girl anime stuff.
So a higher Dim and more training time might help but like
>>8639025 said this also means you have to adjust training time. I'm no expert but I imagine having a higher Dim means it'll take longer for the network to learn since it's larger. If you were using that config maybe bump the Dim size up from 32 to 48 or 64 and increase the alpha to maintain the same ratio. Then maybe increase the LR by 50-100%, or just let it run for more steps.
>>8638064Well how hard is it to pilot *well*?
What ever happened to right clicking for metadata?
>>8639080They got rid of it after 4chan got hacked
>>8639080many people stopped sharing their metadata
>>8639083Yeah but the feature itself is gone post-hack
>>8639084nah still works
did you uninstall the userscript because you thought the site was dead and gone?
>>8639085NTA but there's a script?
>>8639088>>8639089oldfag who's been away for awhile
last thing I remember was that rouhei model people were hyping for no reason
>>8639023> Pic for comparison: https://files.catbox.moe/4usl1d.jpgIs lora left or right?
>>8639025Just keep the ratio at 1.
>>8639065> IIRC the network's Dimension is basically how many neurons you're adding to the modelIt does not add, it changes values of the model weights by applying small matrices. The bigger (dim) the matrices, the more accurate resulting weights are, but it does not guarantee the quality. Alpha here is how "strong" the applied effect of the matrices. And training time should be the same.
>>8639023I also said it only matters for characters, not styles. But sure, give it a try.
You didn't provide metadata, but it looks more like you're testing in on WAI or another shitmerge. Those apply shiny skin on top of everything.
>>8639127>Is lora left or right?Clearly right, with fucked fingers and less details.
>>8639142and the bright skin shine he was complaining about
>>8639023You using a trigger tag? Gotta be using a trigger tag. Anyway nah, higher DIM will most probably not help you.
I personally rejected like 20% of what I was trying to bake because it didn't really work.
>>8639023>Magion02Oh and that style looks pretty sloppy and inconsistent too. Did you at least scrape kemono for higher res pics? The danbooru ones are lowres af.
>>8638587>>8638900Done, his colored artworks are very scarce since most of his work are greyscale doujinshi so I can collected only 155 images that I think they're good and not dupes. I cropped and removed text/signature/logo on some artwork tho, you can see them as a png files. feel free to change anything if you think it's better including images.
https://files.catbox.moe/2w6dcl.7z
>>8639186Is stuff like https://files.catbox.moe/mcwkik.png watercolor and traditional media or what? I'm not familiar with the guy. I should probably tag it for less leakage.
>>8639192I'm assuming it's traditional media with marker (medium) according to danbooru
>>8639196Are you the OP with the pic? I'll tag it as marker if so.
>>8639192https://www.youtube.com/watch?v=N52WmVl3xRQ
>>8639203one of my favorite artists
>>8639207for me it's aka6
>>8639127>Just keep the ratio at 1No it isn't ratio based at all. Alpha has an affect on LR based on what the numbers are, regardless of dim. Most people here do some sort of calculation and even then it's not accurate and requires empiric testing on a dataset they're confident with.
>>8639259It literally IS ratio based.
>>8639259https://rentry.org/59xed3#network-alpha
>>8639254>all those missing booru postsuh.. BRO?
>>8639266I'm 99% certain we're talking about different things. I'm saying you obviously keep the ratio the same but that doesn't matter because your settings are shit if you don't change your LR. Finding the correct LR for the new alpha is more trouble than it's worth.
>>8639268
>>8639259And if you want the source from you know, the author of the damn scripts
https://github.com/bmaltais/kohya_ss/wiki/LoRA-training-parameters#network-alpha
mfers just talking disinformation for no reason
>>8639274If you disagree with
>>8639273 then you're retarded and have never touched the settings in question.
>>8639271He's one of the strongest artist on noob. Got like 700 pics on e621, and the style is very different from the model baseline.
>it's not x
>it literally is here's official source
>schizobabble
every time
>>8639186Looks like it's working?
https://files.catbox.moe/lw590f.safetensors
>>8639295except it added extra toes
and still got that skin shine as if we were back on Pony
oh yeah nta
>>8639326autismmix.. my beloved
dont forget to lather grease and tire shine on your gens bros
>>8639326you might want to take a look at the dataset bro
>>8639367oh yeah, never mind
I honestly did, but only the first colored pic on danbooru which didn't have any of it https://danbooru.donmai.us/posts/8757513
i've rediscovered the joy of genning after baking those artists lmao
>>8639404that sounds nice
>>8639295requested anon here, after some experiments looks like style works better with 291h compare to 102d custom (I only have these both) with euler ancestral cfg++ and sgm uniform, for feet part I have no idea what causing this I also notice that sometimes navel are a bit weird but it's not a big deal for me I just inpaint them later. Working great than I expected so far, tummy fold and ribs details are my favorite thanks to you. Next time I maybe request for greyscale version, if you would like to of course.
>>8639491Hell, I can bake it, but I bet this'd work too with prompting.
Chat, I'll give you around 4 hours to create a new thread or I'll bake it again
>>8639532Can't we just merge back into /hdg/? I'm tired of having to check two places.
>>8639023gave it a try with my settings
https://files.catbox.moe/9q61bi.png
the problem is that this artist has like 8 different styles
![GAr6U61WcAAVxFC[1]](/images/h/1750437216157311.jpg)
![[nsfw], artist_{masterpiece artwork vibrant, colorful, karacak, Syuro, miaru, ya s-3706289151](/images/h/1750531456506390.png)
