Previous thread >>8986916 The [dead] local Stable Diffusion realistic thread>Uncensored Flux modelChroma: https://civitai.com/models/1330309/chroma>Wan video modelshttps://civitai.com/models/1329096/wan-video-21https://civitai.com/models/1817671/wan-video-22>Related threads>>>/aco/futasdg>>>/b/ai>>>/gif/vdg>>>/g/sdg>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/trash/sdg
>>9015395>https://civitai.com/models/1330309/chromaOh, I missed these threads, there is an uncensored flux now? I'll have to try it out
reminder that even VRAMlets can use video models and loras to make single images.
>>9015395Thanks for the bake, it's been too long.>>9015452Been out a couple months now, it's trained by furries though and afaik even those aren't too happy with it. Last time someone tried it here he said you basically have to AI-generate a prompt that fills up the token limit otherwise you get bad quality. pic unrelated
>>9015654Fixed the broken sword
is there any decent civitai model for photorealistic porn that follows prompts? SeeDream is by far the only model I've found that is good for photorealism that actually closely follows prompts and poses, and doesn't mix up attributes between multiple characters.However, for some reason Seedream doesn't do well with NSFW generations. It doesn't block them or anything but it's just really really bad at them, to the point where I have to assume the creator intentionally didn't want that model used for porn. It's like Seedream has never seen a penis and doesn't know what a blowjob is.Also why is Seedream so bad at NSFW? Most other models will generate NSFW, they just suck at everything else (following prompts, faces, realism, etc). Seedream is great at everything else and then for whatever reason is terrible for hardcore stuff.
>>9016460chroma maybe, if you're ok with the rnd
>>9016460Short answer is try Chroma https://civitai.com/models/1330309/chroma or dumber checkpoints but specifically focused on nsfw, like Lustify https://civitai.com/models/573152/lustify-sdxl-nsfw-checkpointOr browse Civitai for pics you like then check the model and prompt used, go from there.>follows promptsThat depends on what your prompts look like. True natural language understanding is out of reach of anyone other than the huge online services, chatgpt, cortana, gemini, grok, dall-e, sora, etc. The main issue with NSFW in photorealistic models is that the models are not smart enough and there is no standardization in dataset tagging. For best results you need to know how the creator tagged his images in training, and then use similar words/phrases/syntax. You can get a limited form of that on Civitai if they provide some preview images with explicit hardcore, or from others' pics in the gallery if they randomly figured out how to do a specific thing. afaik most checkpoints capable of hardcore beyond the basic pov cowgirl/missionary are at least partially based on SDXL anime models, and as such require you to prompt using Danbooru tags, eg. instead of "girl sucking dick" you'd write "1girl, 1boy, from side, upper body, fellatio" each referring to a specific tag on the wiki https://danbooru.donmai.us/wiki_pages/fellatio. This standardization means the model can learn the concept easier and doesn't get confused between sucking cock, blowing a guy, giving blowjob, performing fellatio, and a dozen other ways to phrase even just this one thing. The downside here is the underlying anime knowledge often leaks through and can give you uncanny valley looking stuff. Think the most well-known one in this category is PonyRealism https://civitai.com/models/372465/pony-realism
>>9016460I don't know exactly what you are looking for but for realism I use https://civitai.com/models/443821 I guess give it a try?
Helping a blacked guy is a -1 INT move.
>>9016600Eh, I help futa fags on /d/ too. There's enough content online now to stay angry at every day for the rest of your life, I don't think removing one creator of a hated fetish would make a difference there, nor a hundred. But he may pass it on to someone who ends up creating something more uncommon and more palatable.
>>9016605>I don't think removing one creator of a hated fetish would make a differenceyeah, no one has ever made an entire board to contain the autistic energies of one group of people.
>>9016596yeah, I've tried that model and it's maybe the 2nd best behind seedream but it's a distant 2nd (although it's a lot cheaper). The realism isn't quite the same, it's close but it's still clearly CGI whereas Seedream really looks like real people. But the biggest difference is adhering to prompts. Seedream is just far better than anything else in strict adherence to prompts (at least for non-NSFW). Especially with multiple characters, Seedream can adhere to individual prompts with 6 different people while every other model will mix attributes between the characters or not even have some of the characters. Seedream is also good at prompts for specific facial expressions and poses that other models aren't good at. Idk, Seedream is basically perfect except that for whatever reason it's terrible if you try making porn with it.Pic-related was the exact same prompt as >>9016460. I did another generation that was more accurate for the blowjob but there still wasn't a second woman and it still didn't really seem photorealistic like Seedream. The seedream got everything right except for the blowjob itself.>>9016539>>9016554I just tried Chroma. It's even more expensive than Seedream (50 buzz vs 40 per image) and it won't generate, it looks like it's at least temporarily down.>Unstable Resources>The following resources are experiencing a high generation failure rate, possibly due to temporary generator instability. This usually resolves with time and does not require action from you. This notice will be removed once performance stabilizes.and then when I try to generate it doesn't and gives this:>This is taking longer than usual. Don't want to wait? Cancel this job to get refunded for any undelivered images. If we aren't done by 1:23am we'll refund you automatically.so, idk, maybe it's temporary or maybe it's not supported anymore.
>>9016638Is it because Civitai doesn't allow you to use blue buzz to generate NSFW anymore?https://civitai.com/articles/20211
>>9016641damn. That just started today apparently so I don't think that's it but wtf. I hate how much companies are caving to censorship demands from banks
yea sucks it was convinient to just onsite create luckily I set up my local stuff before this was even announced but still sucks to not be able to just quick gen something without firing up the local
>>9012432Wow these look incredibly clean for AI images.I wonder what he's doing other than using a fat GPU.
>>9015719>>9015722I should make some Nilin sloppa>>9016460Seadream, just like many other models straight up doesn't know what's there between legs. It simply can't do these, no matter what you prompt.As for realistic porn there is:Some SDXL finetunes (mediocre, but can work, and currently the best bet if you want maximu realism)Realism mixes of Pony and Illustrious (better porn and even better for niche fetishes, but the realism is weaker than SDXL)Flux with loras (Bad since base flux has zero clue about genitalia and being distilled doesn't take training new concepts too well. Can get results with some luck though)Chroma is like Flux but it actually knows NSFW. The model is kinda mindbroken and schizo about anatomy at the moment so requires luck to get good results, at least until someone makes a good finetune.Wan 2.2 t2v with some NSFW lora can also make realistic images with image gen (single frame) mode. The base model lacks NSFW knowledge so the model is iffy when trying to get around that with LORAs like Flux.
>>9015395
>>9016867>>9016868>>9016866wow prompt?
Does anyone know how to inpaint with flux without using the "Flux Realistic" sampling method? This is for forged, but every other sampling method just makes it blur. The reason being is one of the Forged updates actually removed those sampling options and I had to revert to an older version, so I can't really update it without losing those.
>>9018349just use euler a and whatever scheduler. you need to inpaint at higher resolutions for it to look good.
>>9019050the thumbnail looks like something else
>>9019598...someone made a lora for that.
>>9020219What, for misleading thumbnails? I don't think scat needs a lora at least.
idk if i can even upload images on this browser
Has a girl ever Goatse for you?
Anyone have a comfyui guide for rocm/amd for arch?
>>9023020I am an arch user.I don't know any arch specific think you need to do. The usual git clone create venv activate that pip install -r requirements.txt works. (Though I prefer to use docker for extra security, I don't trust extension code that I don't take time to review)I don't know shit about rocm so can't comment on that.
>>9023396Catbox?I would like to know how you got this style.
>>9023398It's XL and a doll lora [https://litter.catbox.moe/oj83xt.safetensors] and then img2img with Chroma.
>>9023485Thanks anon.Out of curiosity may I also ask what de-noising do you use for Chroma img2img?
>>9023490I use a low step FP8 version of Chroma> https://civitai.com/models/1330309/chroma#quantization-options12-16 steps, Euler A. It might give worse results than regular Chroma, I haven't compared them.
>>9023506Thanks but I was specifically curious about denoising strength, do you use something low like 0.3 or higher values?
>>9023511It needs to be higher if you want it to fix XL mistakes, around 0.5 or more. Lower if you want to keep the style.
>>9023524That sounds about what I would expect but still useful to get confirmation. Thanks again.
Is there a good model to generate images featuring bondage?
>>9024090I don't really get this fetish, but at first glance it seems to work the same as...well almost everything. You take a model based on noobAI that actually understands what you're asking for, and then refine or img2img with a model that's utterly braindead but at least kind of looks like realism and not uncanny valley pseudo-anime. In this attempt they were Sambase and Lustify, respectively. There's probably better options for both by now but you get the idea.also, fuck fingers
>>9024650>I don't really get this fetishI didn't until I saw this video of a guy dressed up as a pirate on a pirate ship stage tying up his svelte lackey who had black hair and no panties in increasingly humiliating ways in front of an audience.
>>9024090bondage is tough as a one-shot because of inherent limitations in how these models determine the relationships between different objects. A lot of bondage tags fuck up hands or the overall composition and you have to get kinda creative. Chroma v44 or v48 are ok if I recall correctly>>9024840i have had this fetish since before i knew what sex was. I remember being captivated by this picture in a book about American Indians where there are some guys getting suspended from a ceiling, and this other book about pirates which had illustrations of a line of captives chained together (in their underwear IIRC, which was wild for a kids book if you think about it)
Question. One way I figured out how to fix the "ai slop" is by genning pictures with larger dimensions but are there other ways to fix bad eyes/hands?
>>9025203inpainting and/or using a better model, Flux/Chroma/Qwen like >>9023506and loras might increase model errors.
>>9023306oh man the future is weird
>>9024090This is rare but interesting to regularly attempt. I'm not sure if this could directly answer your question>good model to generate images featuring bondageI was on this for a slight while, for several weeks. If you prefer prompt adherence, Illustrious models have been effective for me. Also knowing the basics of danbooru tags, prompts - i.e. bondage, shibari, bound wrists, arms behind back, arms up, and you could try and see how your first gen/base gen turns out. For rarer bondage positions, you might need to find loras. The typical positions that you can prompt immediately seems like arms behind back, that gives usually a boxtie. Also from my experience, the ropes tend to change and add up when you gen several times with the same seed.But if your goal is visuals as well as realism, I'm unable to advise much, but like yes you can try the models mentioned by the earlier anons, img2img using a realistic model after getting a decent base gen perhaps.I won't attach an image here but could include catboxes as this is a semi realistic model I used, that I think isn't suitable for this thread but liked it for its prompt adherence.https://files.catbox.moe/j21uxz.pngBase genhttps://files.catbox.moe/9s1hin.png Adetailer fixed, especially the eyes and nippleshttps://files.catbox.moe/te39j9.pngBase genhttps://files.catbox.moe/fwvdka.pngAdetailer fixed for face.>>9024840Not thread related but there is a small community with the same interest https://boards.4chan.org/aco/catalog#s=female%20bondageJust only a precaution that a few cavemen/traditional art fans are against AI there.>I didn't untilStrangely there were 90s cartoons of a girl character constantly captured and always a damsel in distress/DID. That piqued my interest but couldn't understand it when younger. Only later when films like Fifty Shades made this fetish more mainstream that I was finding communities and stumbled on that cartoon female bondage thread ultimately.
>>9025203inpainting, and Adetailer which is almost the same thing. They perform a localized crop+upscale of the selected area, so you don't have to blow up the entire image to 4K. Also lets you re-generate the area as many times as you need, while keeping the rest intact. You can also draw in a missing finger by hand, then inpaint to hide the seams or fix shading.Flux/Chroma have a better resolution VAE, so they do not suffer as much from low resolution issues in the first place.
>>9025231>>9025762I have been practicing with inpainting. It helps with basics but other times it's a pain. I think overall it's been a great help but you still got to make the choice for a manual edit or not. Thanks for the help.
>>9025400This is one of the best I've ever seen
>>9028344imo kind of defeats the point of it being AI if she just looks like a normal girl, plenty of normal girl vids out there alreadybut yeah the quality is nice
>>9028353>the purpose of AI is hyper and futa Kill yourself
>>9028482this is the non-futa thread
resorting to chinese so the model understands better.
ai is amazing
does anyone have the workflow json files from this page?https://rentry.org/wan21kjguidethe links are dead
>>9029362catbox where the workflow files are at is down currently, the site has been down several times recently.
So with inpainting, what are ways to inpaint entire objects in?
>>9029657Well, either you go very high denoise, I'm talking 0.8-1.0, replace your whole prompt with only the object and what's in the inpaint area and its immediate surroundings, optionally use the base SDXL inpainting model (for sfw objects) or foocus inpaint patch (if your model is compatible) or an inpainting controlnet, to help prevent the hard seams and loss of context that come from using very high denoise.Or you draw/photoshop in the object you want, then run a lower denoise inpaint over it to make it blend better. Here your denoise level depends on how close your edit is color-wise to the desired result, there's usually a breakpoint where it suddenly turns into what you want. Go too high and your object will not be recognized. You should also prompt for it.I prefer the latter, although it's a bit easier in anime than realism. also yeah i missed the fucking thumb
>>9029657Flux Kontext/Qwen edit can modify images, idk how well it works with art styles or non-normie subject matter.Otherwise you need to have a base image suggesting the object, and the model needs to have knowledge of it, and you have to use the right tag (or lora) or no tags at all.
>>9027496Alyssa is 38 years old this year... eternally youthful hag.
>>9029684Oh right, I forgot we're not stuck with just SDXL anymore.
The RAM kit I bought is now nearly 200% more expensive than few months ago.covid/cryptoslop/AI, it never stops.
>>9028987I guess no one's uploaded a lora of her.https://mega.nz/folder/F2gDBQrZ#McuQdGg_CrvVYRJAId2lCQ
Pretty sad that the indians won't make their own blacked/cuck thread, so normal humans can have their own.
>>9028344thank you
>>9028353this is a good point and it's why i get kinda sick of the 1girls, but that said, it is still possible to gen some very attractive pictures, and from there make them flash you/have random orgasms
>>9029311>>9029321based LeanBeefPatty connoisseurcatbox please?
>>9031336can't tell if cum on body is overpriced or everything else is too cheap
>>9031359lol, yeah, while impressive how well chroma can handle text, its still fairly inconsistent
>>9031666peefag here's your ai /ic/ post. her head is too big and her bones is all fucked up. this is why we don't use realism merges.
>>9031718No way, I make almost same picture by accident.Now I feel like I make plagiarism
>>9032184I'm guessing Frieren with her hair down?
>>9032201It's the back view version of this random gen with a 2D lora forced on the wrong model.So it just got the hair a bit wrong.
it's fun, exploring chroma's nightmare waifu capabilities
>>9032952
>>9032952das some crazy shit
>>9015395SH Homecoming contained multitudes.
you know who you are
>>9015470Wan? But it's slow as hell. How did you make 3d look?
>>9015403Fuck yeah. More?
finallyi can gen buildings i'd want to fuck
>>9034407Yes, I'm the guy who loves everything about these except the jizz
>>9034858This looks like one of those old MAD Magazine gags where an architect mistakenly leaves a porno mag in his blueprints before sending them to the contractor
>>9034858What program is this?
>>9034990krita with Acly's ai-diffusion plugin. it lets you choose between any of the major models and then downloads and installs everything for you
What's the difference between illustratious and pony when it comes to models anyway? I noticed those two are the most popular.>>9029683>>9029684>Or you draw/photoshop in the object you want, then run a lower denoise inpaint over it to make it blend betterI see, that was the issue I was having. That makes sense. Thanks for the tip.
>>9035018Pony came first but it's only really useful for 2D characters and uses a scoring system that diffusion in general deprecated as a concept a few years back. Illustrious is more versatile, less autistic. Both are derivatives of SDXL.
>>9032840>>9032952It's close to Halloween. We should definitely have more monstrous girls>>9035031Oh absolutely. I chucked Pony ages ago, and haven't looked back.
>>9035018>What's the difference between illustratious and ponyVery little, in the realism space. They are both cartoon/anime models, and by the time you push them far enough into this visual style there isn't much left of what they were.In general. Pony had a dataset of one part each anime, furry, pony and cartoon, or more specifically danbooru, e621, derpiboori and rule34. Only, the rule34 dataset is notoriously poorly tagged, and nobody cares about cartoon horses, so that's half of it doing more harm than good. illustrious is newer and had a more focused dataset that included only danbooru, and a lot more images total. So it has better knowledge of characters, styles and concepts, but listens primarily to tags from the danbooru tag wiki rather than natural language prompts. There's also noobAI which continued training illustrious and reintroduced e621, for more /d/ and furry adjacent concepts as well as deeper danbooru knowledge with the additional training.But how any of those bases translate into a photorealistic checkpoint depends more on the person re-training it with photographs than on what the model was originally.
>>9035168NTA, but thanks for the great explanation.
>>9034732>>9034736can you post prompt for one of these?>>9034919fair
>>9035419sure :)for >>9034732 :an amateur close up photo of a lineup of wet nude lower halves of two hot young women sticking out of a low makeshift wooden wall at waist height at a remote hiking trail with a scenic view. you can barely see the backs of their heads sticking out on the other side of the wall. their legs are spread far apart, standing with their feet on the ground. their backs are arched. some of them have light pubic hair. there is a official sign nearby, saying "Public relieve station. Please clean after every use" with a waterhose coming out of the wall underneath, rolled up on a hose holder on the walland for >>9034736 :an amateur close up photo of the sweaty nude lower halve and lower arched back of a hot young woman affixed and sticking out of a low makeshift wooden wall at waist height at a remote hiking trail with a scenic view of mount fuji. her legs are spread far apart, standing with their feet on the ground. there is an official sign nearby, saying "Public relieve station. Please clean after every use" with a waterhose coming out of the wall underneath, rolled up on a hose holder on the wall
>>9035425thanks
is "releive station" a britishism lol we say relief in ny
>>9035425this is a really good concept, excellent work
>>9035993thanks bro nice to see your additions to it too
chroma? is that what I need to go download now? i'm a bit out of the loop. last thing I downloaded was a bunch of illustrious models.
>>9036186yeah, illustrious ( or one of it's offshoots like noobai or waifuillustrious ) for hentaiand chroma for 3DPD. i've used chroma-flash-HD for pic related
>>9036192does it load up fine in forge or is there some special thing I have to do? what are some chroma merges you guys would recommend?
>>9036186chroma is pretty cool. it's uncensored and follows prompts pretty closely. i've been using it for a while as it's been developed. hands and feet can get weird sometimes. sometimes you also get a weird paper-doll effect, you'll know it if you see it. when you get a bad result sometimes the best option is to do another gen, sometimes you get a bad seed.if it's your first time using chroma, you may fall into a genning fugue, make sure you have lots of water nearby lol
>>9036245boomer prompting feels so cringe after like two years of using anime modelsbut at least it doesn't make everyone look like an airbrushed k-pop star
>>9035993>>9036014seconded
>>9036403Jesus Christ
>>9036339Boomer prompting was cringe because it exposed how little prompt tokens/tags mattered on old models except for some core ones.Hence people who type word salads and get coherent sl0p.
>>9036410Wow!!! Goddamn looove it
And now for something old school, whether or not anyone wants to modify or animate it, idc
Happy Halloween! Post spoopy gens!
Why is it that all the models on Civitai are 6gigs? I know nothing about anything but is something like a 30 gig model that has a more wide breadth of accurate images it can create a thing?
>>9037075All SDXL-based checkpoints are 6 gigs; Flux-based checkpoints, include Chroma, are 17 gigs. A 30GB model could only be loaded on a 5090 (32GB VRAM) or better video card, and even then the 2GB left over might not be enough for LoRAs, upscalers, feature recognition and so forth.
>>9037105i seeee, ok makes sense.
>>9036245My chroma must be broken somehow then. I've never gotten anything out if it other than blurry shit in a pseudo-anime style.>A photo of a blonde soldier girl with gigantic breasts. She is holding an assault rifle, kneeling on one knee in a storm, rain, wearing military camo with exposed cleavage, black gloves and kneepads.
>>9037105Hugely useful to know when specing out a graphics card, thanks
>>9036245Humor me, can Chroma depict a thumbs-down gesture? Baseline Flux appears to have literally blocked this no matter how you describe it.
>>9037170I've been unable to do it, tried lots of different phrasings.
>>9037227Looks like it's time to train a LoRA!
>>9037227Thanks for the push, making my shitty-ass prompt work. I figured it out, was using a workflow for base Flux, with dual clip loader, no tokenizer padding, etc. Didn't realize Lodestone changed the architecture so much.Still damn blurry tho
>>9037034some more chroma nightmare waifus coming right upprompt: a grainy amateur photo of a nightmarish hot female creature in the woods at night
>>9037533
>>9037562
hey comfybros, please save your images with stealth metadata node, so we can all see your workflows, as 4chan removes regular metadatathaks !
>>9038360Same for webui/forge. Or use the 4chanx userscript to upload to catbox simultaneously with one click. Somebody vandalized the OP, it used to be there.I don't include it unless someone asks, stuff's pretty much all img2img or inpainted so metadata is not useful on its own.
>>9028038>>9028040>>9031156>>9033300>>9035419>>9035494more of these trad beauties please!
>>9038809i am out of tensor credits plz specify specific race and phenotype and i'll get to it when i get to it
>>9039266hot
>>9036633this is very true, either a token mattered or it didn't and there was nothing you could really do about it except hope a synonym would work better>>9038809
>>9039423>>9039285Imagine living in a beauty bottomless society with women like that.
>>9039264this >>9039264 phenotype, race: central/ northern european (like before), preferred blonde, brunette ok, ginger is also fine by me >>9039285, though if ginger preferably thin/skinny, brunette and blonde can also be "homely"/ curvyyour efforts are much appreciated, these gens are fantastic, thank you
I noticed some days that the gens work better then others and some days the art style is completely different even when using a style lora
>>9040070sounds retardedhave you tried re-running a gen from yesterday? see if you get the same result
still refining, any suggestions?
>>9040307
>>9040307I suggest boobs
>>9040319furthest I've gotten to with gemini, I'd be glad to make it more explicit, but damn, it's a hard puzzle to solve
>>9040393
In your view how long shall we wait for a really good realistic AI model which has been trained on OnlyFans, Reddit, X and Pornhub? (Chroma seems promising but both its terrible anatomy and its low knowledge of genitalia is a bummer; the only realistic models which have impressed me so far are BigLust and EpicRealismXL although they still can't make feet look bearable and they might occasionally look too 3D-like)
>>9040486Most likely people are going to keep training Chroma LoRAs and maybe merge them back into the checkpoint.
Maybe I was just too lazy but I could never get chains to work correctly on a consistent basis in Automatic, Chroma is looking promising
>>9040419>>9040422>>9040427what did you use to get such good looking cum in mouth?
>>9040422catbox?
>>9039928homely/curvy ginger, understood
>>9040883Oh jesus
>>9040960>bracesgood old nightmare fuel on SDXL
>>9040886make bush great again
>>9040983Wasn't he always great?
Does anyone have a link to a pussy rubbing, female orgasm, or machine fucking wan lora on tensor? that site's search function is unusable
>>9041405there are loras for all of that for wan 2.1. don't know about 2.2i got them from civitAI, but they may be taken down by now. so also try civitAI archive
we sliding
new thread>>>9042333>>>9042333>>>9042333