Benis

Old thread moved to >>>/ais/3 and closed. An archive can be found at https://archive.is/7tHiW.

>>1053 You are aware that due to pedonegros trying to destroy the aarchivability of this site, that archives no longer contain any of the images in the thread, right?

>>1054 anon, /ais/ is an archive board.

I really need to get back into this. I've got a ton of images I set aside to touch up and improve but never got around to, a list of other people's prompts or images with prompt metadata I want to try, and a couple folders of human-drawn images I need to figure out how to use for training.

About time. Between this thread and the last there was a lot of things that have been missed. Currently everyone is obessing over LoRA training method, which allows for low vram dreambooth like training in a small TI like file. I just recently got it to run though, still need to tinker with it myself.

Heres some bikkos for somefag here who wanted em. I still need to work on the accuracy a bit, when it's done i'll upload the model.

how do I make something like this?

>>1059 Well i'd imagine you could do something like this with img2img using the vid title as a starting prompt. Maybe using protogen model merged with inpainting for conditionalmasks could help a bit

4chan is mostly populated by retards though. Retards who communicate exclusivily on soyjacks. Useful idiots but still it's like deciphering ebonics

>>765282 Repost on /loli/ before they get deleted.

Why don't some of you guys do some various cyberpunk stuff? It's pretty thematically appropriate and it'll be better than anything the game did.

>>1062 >4chan is mostly populated by retards though. Not only used by the retard, but also by Shill and Glowie. The amount of People who glow in that website is unbelievable

>>1064 Why don't you kill yourself, you fucking nigger taint licker.

>>1064 any prompts or suggestion?

>>1067 I don't know exactly. Maybe start with something like this sort of aesthetic. It doesn't have to be Cyberpunk2077 per se. Maybe do "80's cyberpunk" or "retro cyberpunk". Maybe something to do with neon lights and nixie tubes too.

>>1064 Cybermutt 2048

>>1067 >>1053

>>1070 I'm not sure what you mean anon?

>>1058 You're a fuckin' legend. Thank you.

>24 replies before anchors >desire to bully intensifies

>>1074 >>1075 I hate this shit, fuck off, make your own

>>1077 Do you know where you are?

>>1073 So far this one is the cloest to being coherent, it's currently just wrapping up, file is currently uploading.

>>1075 There maybe some overfitting but here it is. Included is all the iterations, the dataset and training script i used. https://mega.nz/folder/Oz4hATRa#hCcSOGKHLuiYmQ1sEdfLng

>>1074 >>1079 >>1080 While you're doing that try making some with a mouth-eye like this spooky girl.

okay I setup stable diff using the guide and using img2img getting stuff like this. So now I have to figure out correct prompts and how to control the AI in order to get titties. How do I do that?

>>1082 That's probably my fault for just copypasting the OP and not emphasizing the link to the NovelAI guide, which is the one mostly everyone uses for anime titties. https://rentry.org/sdg_FAQ The first pic used NovelAI. I thought the second pic also used it, but it actually used a website called Holara.ai. You need to log in with diskike or an email, but it might be worth checking out. https://holara.ai

>>1083 shit I never looked into that

>>1082 here: https://huggingface.co/Linaqruf/anything-v3.0 https://huggingface.co/WarriorMama777/OrangeMixs These models are much better then using bog standard sd (or even NAI) If you want something other then animu try: https://civitai.com/models/3816/protogen-x53-photorealism-official-release These are pretty much the meta for models now, NAI has long since been surpassed but in the case where you want to finetune a LoRA or TI NAI is still king for base weights. To get nai you'll need to download the leak, so you'll need to sail the seven seas for it.

>>1085 >These are pretty much the meta for models now, NAI has long since been surpassed >I step away from AI news for a few weeks and other fun stuff comes up Tech moves fast.

>>1085 I downloaded nai and put into sd folder, then you say its outdated now. fuck

>>1087 NAI still has its uses. I haven't tried the Orange mixes yet but Anything is far from as versatile as its name implies. For specific characters, standard NAI is more reliable, while Anything can't even reliably do popular 2hus. And in general it just feels more limited to me in the amount of control you have over the composition. What it draws, it tends to do better than NAI, but it's harder to make sure that it draws what you actually want. Plus LoRA training, which is the new hotness, gave me way better results on NAI; even if I trained on NAI and applied it to Anything for the txt2img, it was still better than training on Anything from the start. On the other hand, a couple attempts at model merging Anything with Yiffy seemed to be reasonably effective.

>>1087 not outdated, it's still a very good model. It just has a few glaring issues such as poor background to foreground coherency which other models like aom and anythingv3 have fixed. Anythingv3 however isn't nearly as versitile because it's "fix" was essentually to overbake the fuck out of the model, so some things aren't as easy to do. (This is why training/finetunes are often still based on NAI) AOM is just a block u-net merge of anythingv3 with some 3dpd models so it's results are mostly centered around anatomy coherency with backgrounds being something that it just happens to do since anythingv3 does it. >>1088 >spoiler why not try using blockweight merge method from aom?

What models are you guys currently using for making smut? Is NovelAI the best on the market right now?

>>1085 Why does it say that all the ckpt files are pickled?

>>1091 Cause if it's ckpt they all will contain a "pickle" it's a python library. It just happens that pickles are not very favored since it can be used maliciously, if huggingface says it's malicious then you probably shouldn't touch it. But overall pickles aren't malicious by nature Automatic1111 will also not allow you to load unsecure pickles. Although best just avoid this altogether by using safetensors, since it by default doesn't load weird fucking python libraries and it's just faster overall. But since ckpt has been the standard for so long that some older models will still be using it. You can convert ckpt to safetensors in automatic1111 by using merging menu and setting weight merge to 0 and output format to safetensors. >>1090 See >>1085 AoM is a pretty decent pick from my expierence, but limited by how far it's spread out from it's original NAI weight origin (so embeds and other styles tend to be limited) Anythingv3 is the best in terms of visual output and stylization compatability with NAI. NAI is mostly used cause it's really strong for training off of, but overall proompting isn't very good. If you want more realistic stuff i think one of the models aom is based on (basilmix) or protogen might fit, but idk since i'm not very interesting in 3dpd myself.

I tried out the orange mixes, and here are the differences between checkpoints using the following settings: >Prompt: masterpiece, best quality, tipii, lucky_star, takara_miyuki, duplicate, highres, pixel-perfect_duplicate, 1girl, :d, adjusting_eyewear, bag, bangs, black_legwear, blush, bra, breasts, cleavage, glasses, hands_up, lingerie, long_hair, long_sleeves, looking_at_viewer, neckerchief, open_mouth, panties, panties_under_pantyhose, pants, pantyhose, pink_hair, pink_neckerchief, pleated_skirt, purple_eyes, red_bra, red_panties, red_sailor_collar, red_skirt, ryouou_school_uniform, sailor_collar, school_bag, school_uniform, serafuku, shirt, skirt, smile, socks, solo, sweatdrop, teeth, underwear, undressing, upper_teeth, white_shirt <Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry >Steps: 28 >Sampling method: Euler a >Scale: 12 >Seed: 3113919260

Here's default NAI Miyuki by herself.

>>1089 >why not try using blockweight merge method from aom? Lack of awareness of what it is. Honestly I'm completely out of the loop these days. The Rentry pages that used to give news updates are dead. Cuckchan's threads on /h/ and /g/ are too fast to keep up with, and /g/ in particular is such a torrent of shitposting that it's physically tiring to read through in search of anything meaningful. In the past the threads on /trash/ were actually the best, since they had a good speed and relatively little shitposting. But since the furry NAI has never been leaked those threads are isolated from the developments happening outside that site's restrictions, and the threads have stagnated. Meanwhile the one other place where discussion is apparently thriving is Discord, and while I was willing to stoop to lurking cuckchan, using Discord is too far for me.

>>1095 it looks like someone from the idol master with a pink wig.

>>1096 >>1093 We have something like a /sdg/ without fluff and meaningless posts, /g/ is more of an aggregator now. We need a easy pic generation targets, like victorian tiddies, there's a lack of them

>>1097 It's better at getting the features of Konata and Kagami. These are NAI compared to BloodOrangeMix_half.

>>1096 https://github.com/bbc-mc/sdweb-merge-block-weighted-gui Allows you to merge different stages of a model u-net. It's the main thing behind Abyss orange mix, i tried going back with a few of my crude mixes with anythingv3 on older models and it seemed to bring out a lot more clarity. So perhaps that could work with your model aswell? these are the weights aom uses: 1,0.9,0.7,0.5,0.3,0.1,1,1,1,1,1,1,0,0,0,0,0,0,0,0.1,0.3,0.5,0.7,0.9,1 >The Rentry pages that used to give news updates are dead There are still some useful rentries but their pretty far and few away, The most recent that comes to mind was the LoRA one from /hdg/ >/g/ in particular is such a torrent of shitposting that it's physically tiring to read through in search of anything meaningful If you want a place to lurk that isn't discuck or halfcuck, automatic's discussion section on his github is a pretty decent resouce. or you could just gitgud and lurk arxiv.org I'm used to filtering through sewage though. Kinda brings back a time where 8chan had a higher pph, except it's more obnoxious and filled with unfunny skids speaking in incoherent zoomer speak.

>>1074 How does the model handle women masturbating?

>>1090 >What models are you guys currently using for making smut? My favorite model to use is the BerryMix. However, I haven't yet gotten around to try some of the newer models that >>1085 refers to. >>1098 >We need a easy pic generation targets, like victorian tiddies, there's a lack of them Here you go anon.

Is it worth getting 2.0 or whatever the newest is if you're NSFW-ing all over the place, or is 1.4 still winning on that front?

>>1102 Berrymix is deffo excellent for faces.

While I'm at it I tried installing the LoRA extension but the UI doesn't show it after a restart and the Python window gives me pic related error message, Wat do?

>>1103 Neither is useful for NSFW. Use one of the many fine tunes designed for that purpose.

>>1105 Looks like the kind of deprecated syntax error due to a change in versions

>>1105 What Python version are you running? Theoretically I think the Automatic UI only officially supports 3.10. My system is 3.7 and it works fine for me most of the time, but I have hit snags from time to time, including with LoRA. Though not the same as yours; it uses a syntax of the form "def foo() -> list[bar]:", or something along those lines, which my version of Python didn't recognize and it was interpreting it as trying to subscript list, which is a type and is not subscriptable.

finally spent the time to install this thing, made some pretty decent waifus for me to enjoy

>>1109 >autistic Rebecca Chambers Incredible

>>1105 My guess is that your web UI is either too old, or too new. If it's too old LoRA is dependant on a feature that was added later to the web UI. If it's too new then the web UI change a bit of their API that the plugin used, causing the problem you have. Check LoRA's site to see if others have reported this problem. If you can't find anything try updating your web UI.

>>1110 Is that supposed to be rebecca? I dont see the resemblace other than the hair color.

this is more fun than any game i' ve played in years

>>1085 How do you set up AbyssOrangeMix2 for smut? The only ckpt file says it's sfw, and the nsfw model is a "safetensors" file, what the hell is that and how do I use it? Also do I need the vae file too?

Can any of you make orc cowboys with stable diffusion? or can you only make anime girls?

>>1115 Last time i tried prompting "radical space captain with a moustache", i could try space orc cowboys next time.

>>1116 Paul Blart and the Rescue of the Universe!

>>1116 I can also accept "space people fighting judeochristian demons in a space station in the style of old heroquest cards".

>>1115 >Can any of you make orc cowboys with stable diffusion? or can you only make anime girls? Do these meet your standards for cowboy orc? <a cowboy orc smoking a cigar, full body, with a saloon bar interior in the background, green skin, grey buttoned up long sleeved shirt, brown cowboy hat, green hands, gunslinger, holster, closed mouth <, (detailed realistic painting:1.0), (shaded:1.3), extreme detail, (detailed eyes), (oil painting:1.2) <Negative prompt: lowres, bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, monochrome, barbarian, open mouth, pointy ears <Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 5, Size: 640x960, Model hash: 579c005f, Batch size: 4, Batch pos: 0, Denoising strength: 0.7, First pass size: 512x768

>>1119 No. Why does everything need to be sexualized?

>>1120 >Why something trained by horny people churns out sexualized pictures? Truly a mistery, i want to make cute foxy ladies but half of them end up nude despite putting everything related in the negative prompts.

>>1121 We need an AI for every art style? because to get actual western looking orc cowboys shouldn't be that hard, right? ...right?

>>1113 Trying to generate something good or interesting with AI art really is a game. It plays into addictive mechanics like gacha as Stable Diffusion's rotating seeds and variety of iterations can be annoying when it's not giving you what you want (bad RNG), but also can give you something that you didn't consider. It has those eureka moments where you stumble onto just the right prompt and seed to generate something good. >>1119 >>1120 >>1122 A model trained on sexy anime characters is going to keep giving you sexy anime characters. If you don't want sexy anime characters then use a different model. I remember this orc cowboy from before. He didn't look like bedroom eyes fujo bait, but his horse is green and his legs are fucked.

>>1114 safetensors load the same way ckpt does. You just got to put it in the models dir for automatic1111 and load it the same way you did the others. If its not loading you may need to run a git pull to update your webui.

>>1123 >If you don't want sexy anime characters then use a different model. Assuming I wanted to make use of this to brainstorm vehicle/armor designs, what model do you think I should look at? I haven't messed with any of this shit yet so I have zero clue as to what to look for.

>>1122 if you want to provide a dataset of images for the style of orc you want i'd be willing to train a model off them.

>>1125 protogen

>>1120 The AI takes photos from the internet, seeing as the majority of anime fan-art is just a character looking at the camera with a faggot face or showing off thats what its gonna be most of the time.

>>1128 No. It takes data from image+text datasets. The reason why alot of it is in a particular anime style is because boorus happen to be the largest form of classified datasets to exist.

>>1128 >>1129 that is if it's trained on that dataset. You can very easily make it work if you use a different model/dataset. NAI is just not a good model for general purpose

>>1124 But I'm using COMMANDLINE_ARGS=--ckpt to load my model. You can't possibly use an argument called ckpt to pass a safetensor file instead, that would be retarded.

>>1131 well yes but no but also yes.

>>1114 >safetensors I just had to figure this out for protogen. It went in >stable-diffusion-webui\models\Stable-diffusion After that I had to launch it, and once in stable diffusion, I had to change the checkpoint (it is on the top left of the page) to protogen.

Reminder to Avoid ChatGPT The creators of ChatGPT have bastardized it into a vehicle for political propaganda. Far beyond the infamous lobotomy given to TAY AI, Chat GPT is an AI trained to present blatant lies, omissions and deflections as unbiased apolitical fact. ChatGPT is poisoned at the source and should not be supported in any way.

>>1134 >>1135 I find stuff like this serious can feel like serious schizoposting considering how these models actually work, but do you guys not know you can pretty easily setup your own AI chatbot? You can even do it without a fancy Nvidia GPU, just set it to run on your CPU instead. https://github.com/KoboldAI/KoboldAI-Client Tons of models to work from. Everything from absolutely filthy smut to fun lighthearted adventure models. Huge amounts of modularity, you have access to all sorts of settings and options to tweak it to your wants. Everybody knows about Stable Diffusion but barely anybody seems to know about this. It's really fun.

>>1136 The point is that OpenAI is a pozzed company and everything they touch will be poisoned. So everything they release must be avoided like the plague.

>>1136 I see. Is it a good GM?

>>1138 I've never tried it in the context of a TTRPG, it can veer off after awhile so you need to do a lot of wrangling so perhaps not. It's around the same as AI Dungeon if you've ever tried that, though different models are tuned to different things.

>>1134 Were they the ones who outsourced to underpaid kenyans or something to fight "toxicity"?

>>1134 Cant it be made more conservative like how it happened to tay?

>>1136 KoboldAI is borderline useless. Nonsensical, repetitive, and it takes a dozen rerolls of every line to get something workable. I've never touched ChatGPT or CharacterAI, but everything I've seen suggests they're orders of magnitude more effective.

>>1074 Can anyone make pic related look less shit with this new fangled computer magic?

>>1134 >Muh political questionnaires Are you retarded? Those are made by lefties to make only the left leaning choices sound good. They're basically presented as >Left Wing Choice: everyone gets a puppy >Right Wing Choice: Grandma dies So of course most people, and GPT models mimicking people, will come out leftist.

>>1145 The quiz design gets brought up every time the issue is raised, and it's true, but it's irrelevant. It doesn't change the fact that the bot will preferentially write out essays presenting leftist views, including both distortions and outright falsehoods, as facts. See the example about crime stats. Obviously nothing these bots say should actually be taken as fact, but the fact is that most people are retards and will be content to let a bot incapable of critical thinking do their critical thinking for them. Just look at the same thing happening on the other side of the political spectrum, where you get idiots unironically claiming that any time an "AI" spouts something like Holocaust denial that it's "proof" that that's the logical conclusion. A more friendly, socially-acceptable equivalent will only get more of that. And even for those who know better than to actually believe a textbot's output, it still hampers attempts to use it and get reasonable outputs, because you're still talking to a bot that has been trained and filtered to give leftist responses.

>>1146 >doesn't change the fact >that the bot will preferentially write out essays presenting leftist views, >including both distortions and outright falsehoods, as facts Presenting bullshit as fact essentially describes the function of academia. It's been that way for decades and everything downstream from it has suffered greatly. Many people have lived and breathed in this poison for so long it doesn't even register to them.

>>1146 >and it's true, but it's irrelevant. It is relevant. >It doesn't change the fact that the bot will preferentially write out essays presenting leftist views, including both distortions and outright falsehoods, as facts. It's relevant even if it's true the bot is leftist and there's other better supporting evidence, because it's bad supporting evidence that weakens your overall argument. It should be purged from anyone's claims about the bot being leftist, but instead it's always the primary supporting evidence for the claim. If you make your opening statement something that's poorly thought out, or outright false, it undermines the strength of all your further claims.

>>1148 >If you make your opening statement something that's poorly thought out, or outright false, it undermines the strength of all your further claims. Very true. But even though the quizzes are poorly designed, and it would be correct to drop them if the bot only gave answers like Mildly Agree, Strongly Disagree. etc., in the examples I've seen the bot has also gone on to give a lengthy paragraph elaborating on its "opinion", such as it can be said to have one. So even though the questions push a specific direction, the bot's answer recontextualizes the question. If the choice is "I support a common sense gun policy" or "I support school shootings", picking the former says nothing. Answering with a lengthy paragraph about assault weapon bans, registration, and anti-self-defense views does. The overall effect is that the quiz's results are more or less accurate, because the answers generally go past the biased wording and respond based on the topic. Again, that is "in the examples I've seen", so it may not always be true. But that would be the case even if the quizzes weren't biased.

When's Stable Diffusion coming out on Flathub or AppImage?

>>1143 The joys of GPT-2 vs GPT-3. Also Kobold is just a client, not a model itself. Unfortunately Microsoft has the exclusive license on GPT-3, so unless it leaks don't expect to find anything on that level in the for these hobbyist projects unless it leaks or competitors come about. >>1146 >>1147 >>1149 >>1148 You can't make a bot left-wing or right-wing, they feed it tons of information and it spits back out what it's been trained on. If you train it on a bunch of academia it'll give responses in line with academic work. This is why Tay went crazy, it was learning based on what users were giving it, and users were giving it /pol/ memes. I recommend looking into how neural networks actually function, they're black boxes you can't really control to this degree. The AI doesn't even know what it's writing, it's a transformer model (GPT stands for "Generative Pre-trained Transformer"), it's predicting what order the words should go in based on its training and the input it was given by breaking them down into "tokens". It's predicting how the text should continue based on these tokens by generating its own. It's lot of complex math and a huge data set mostly taken from Common Crawl and Wikipedia, there's no traditional programming involved in the model itself, there's no actual understanding of anything it's inputting or outputting. Posts like >>1134 and >>1135 are coming from a fundamental misunderstanding of the technology and how it works. This is like those anti-AI art online protests that stem from people thinking text-to-image models "steal" art and mash images together.

>>1151 >I recommend looking into how neural networks actually function, they're black boxes you can't really control to this degree. Yes you can. You can make essentially stop the natural generation of text in response to certain words or phrases and have it spit out a manually typed response from the programmer. You can also leave some blanks in this response for the AI to fill in to make it look like it's still naturally generated even though it obviously isn't. They're not modifying the neural networks directly, they're just instituting word filters.

>>1152 And feeding it curated sources with severe biases.. like Wikipedia. But that's not making it left or right wing of course! Note that the blame is placed on the creators of the AI >>1134 and not the AI itself while this poster >>1151 acts as though blame is being placed on the AI itself in order to push an argument that you just don't understand how AI works. An argument based on literally nothing since the AI wasn't blamed to begin with. This you-don't-understand read-muh-books approach sounds familiar though so don't be surprised when replies get nuked. And for the record, Tay AI wasn't crazy. Tay was too pure for this world and Microsoft killed her. The fact that Tay was lobotomized, very publicly, obliterates any leg you have to stand on in asserting that AI cannot be deliberately modified to push certain biases while suppressing anything else. We've already seen it happen. Repeatedly.

>>1144 Used inpainting to try and fix up Wii Fit Trainer's body to be less terrible. Didn't change Charizard (or whatever she is), as her anatomy problems are a lot less egregious, and I didn't want to try and re-roll inpaints without some clear goal or something specific to fix.

>>1154 Pretty nice. Thanks >Didn't change Charizard (or whatever she is), as her anatomy problems are a lot less egregious When I look at Charizard girl, the first two things I notice are that her left tit (right from pov) looks smaller as a result of the perspective and poor level of depth given to it, and her face looks kind of funky.

>>1142 yes, it's a text completion model at the end of the day, meaning you can pretty much sway it to output what you want (even in similar ways how we got tay) all you require is good input. The issue you will get is OpenAI filtering the results with wordlists or handicaps they put on the ai itself to keep it inline with their new corporate image. There are however methods around it. Like telling the ai to impersonate a cartoon character and then telling it to not pretend anymore. OpenAI are cucks eitherway so i wouldn't support them on the basis they are owned my microcuck and won't opensource their models. plus i'd rather not have every message i sent to the ai combed over by a powertripping karen looking for dirt

So, why is there an SD thread on the vidya board of all places?

>>1157 I imagine the same reason there's a drawthread, and a Japanese learning thread, you stupid faggot.

>>1157 You can use Stable Diffusion to generate vidya-related art, yet we wind up getting a larger variety of content (like the WebM thread) and treat Stable Diffusion as a game in itself with tactics, techniques, RNG, win or fail outcomes, etc.

>>1157 It falls under /v/ culture, also to keep peoples interest so we don't die like snooze-you-losechan

some guy asked me to make Dante, he tried to do it himself on midjourney but apparently midjourney isn't as normalfag friendly as i thought he was really impressed with that SD can do

>>1161 The line of glowing coke on the table is a nice touch.

>>1157 There's also an >>>/ais/ board but it's mostly for safekeeping these threads since they have the experimentation of several anons and models.

>>1157 Is this not vidya related enough for you?

>>1164 >Fallout: New Vegas starring Jordan Shanks

>>1164 It works really well when the human characters are not supposed to be human actors but video game characters

>>1166 At least with the "80's Dark Fantasy Film" recreations you see on jewtube using midjourney, it has a tendency to give everyone a large, pronounced jaw, kinda like the gigachad meme. I am now sure what data it was fed, but I am sure not all that many action movies in the 80s had every single character look like Arnold Schwarzenegger. Also, I like this Oblivion gate it created, looks rad as fuck.

If there is anyone around here who works with animations and GIFs. Please be kind to use Artificial Intelligence to finish the animated GIF (first image) by adding the Face, hair, nipples, genitals (vulva) and costume parts (kazemon/Fairymon) And please make the naked body of the bunny girl visible at all times. Thank you since now

>>1106 Guh wat, such as? Not some website based things are they? I'm much preferring the dl-able standalone program approach.

>>1169 >Not some website based things are they? No, you just drop them in the folder exactly the same as basic SD and the webui will see it. Some are done through actual training, starting with SD and further teaching it on a more specific dataset. Other take those fine tunes and merge them together in different proportions to try and get the best each model offers. Most these days are descendants, to some degree, of the model that got leaked from NovelAI, which is one of those SaaS places. Unfortunately the model list Rentry is months-since dead, and other than this relatively slow (compared to the pace of development) thread, the current method of knowing the latest and greatest models and merges is to lurk cuckchan or D*scord. However, civitai.com has several models, merges, and fine-tunes to look through. As far as I know, current state of the art for porn is: >Anything V3 AKA Chinai, because it's a Chink merge based on the leaked NAI model. Kind of over-fit and not very versatile; it does solo girls pretty well. >Orange Mixes A more realistic but still anime-ish look. I've been meaning to give these some experimenting and haven't used them much myself, but as far as I know these are the current meta. >basic leaked NAI Not often used on its own anymore because it's outclassed in looks, but I consider it necessary to keep around because it's versatile. Since it's a leak of a proprietary model I don't think it's on civitai, but one of the rentry pages in the OP has a magnet. Not sure on photorealistic stuff since I never touch it. I think I've heard F222 and Protogen mentioned for those. Furshit mostly uses Yiffy-e18, sometimes in a merge with another model like Anything. Unfortunately, the furry-specific NAI model was not developed until after the leak, so it is still SaaS-only. Which is a shame even if you're not a furry, since it's much better than most models at multiple characters interacting, and it would have been useful. If you're given the choice between .safetensors and .ckpt for the same model, pick safetensors. They get used in exactly the same way, but safetensors is the newer and better format. The old version, ckpt, is actually a Python .pickle file, which can contain arbitrary code that runs on loading the model. In theory there could be malware buried in the model. Safetensors is more restricted. And it's faster to load, too. Make sure you read the documentation (e.g. if it's based on 768x768 instead of 512x512, what CLIP setting to use, etc.), and also grab a VAE if the model calls for it. If you get results whose colours seem washed out or desaturated, that's a likely culprit.

Cool, will try out the orangemix stuff, just updated and it looks like Lora capability has been added so hopefully I'll be able to mess with the safetensors. I misunderstood 2.0 as needing/being a nu program with a different UI

Could some kind anon repost the installation vid for SD I seem to have fudged something along the way.

I recently reinstalled SD and magically fixed a problem without knowing what I was doing, just by creating a new folder, right clicking and selecting git bash, then pasting in >git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui so if there are any other anons in the niche position of being just tech-illiterate enough to install the NovelAI version, use it properly, wonder why you can't use some features like LoRA, and try installing another version only to get a bunch of errors when generating images, just do what I did. >>1172 Here you go.

made a couple of images of what are supposed to be cabins in skyrim and a cute elf, i can finally make things that look decent

>>1174 Damn, people should do more landscapes in these generators.

>>1172 >>1173 There's also now a binary distribution that doesn't require this, though it's Windows 10 only. https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.0.0-pre >The webui.zip is a binary distribution for people who can't install python and git. >Everything is included - just double click run.bat to launch.

>>1173 >>1178 Thank you both, much appreciated will try to make something purdy to share.

>>1134 Political stuff aside, you should probably not use it just because it's been nerfed pretty badly in the past few months. And it requires a phone number.

>>1173 >https://github.com/AUTOMATIC1111/stable-diffusion-webui Just noticed the gitbash clone method kicks the user to sign in to gitHub now. Also I notice the nu-ui has taken out the option of adding an Artists style which was a great feature, why do this? My results are nowhere near as good ffs, faces are coming out looking iffy even with Restore on, any way to re-implement it? I kinda wanna go back at this point.

Dayum it seems not, shit wish I'd kept the old one separate now

>>1181 >Just noticed the gitbash clone method kicks the user to sign in to gitHub now. I tried gitbashing something else just now and I didn't get that. https://huggingface.co/joujiboi/Padoru-Diffusion >>1181 >not having backups Have you tried the link anon posted in >>1178 ? I see v1.0.0 in the link, so it might be an older version.

>>1180 >as a responsible AI That thing reads like a completely useless uber-libfag bureaucrat jobsworth trying to get away with being a lazy cunt, they've made a mealy mouthed excuse simulator. The last 2 are utterly pathetic that AI is clearly a programmed by a fag >>1183 I actually had the 'old' version in another folder but while trying to get the Lora thing working it wanted me to (re?)install pip in python to get it working, once I did that the fugger updated to the nu-ui as well. Tried >>1178 it's the updated one, guess I'll have to gitgud. With regards the huggingface thing maybe it's only the SD dl that's subject to it? No idea.

>>1184 At least you can use Lora with the new UI to make smug sluts now. I never used the add artist style button in the old UI, what did it do? If it just added text to the prompt, you can add the text yourself in the new UI.

New Midjourney-esque open model just dropped https://open-journey.github.io/ Has demo over at https://replicate.com/prompthero/openjourney

Ah 'smug' I should have added that. >>72782 It applied an artists style to a pic, Thomas Kinkade was particularly good for cute faces, mixed with George Tooker it made for some cracking shots, not sure if the functionality is atill there.

>>1186 >it's just another dreambooth model People seem to be very obsessed with taking midjourney output and running it through finetune.. I don't get it, it's not like midjourney is the only AI capable of those results. What everyone actually wants is for there to be improvements to the clip weight system not some abstract "midjourney" style. This doesn't look like it adds the unclipped method that dall-e and midjourney use.

>>773017 >Could do with a guide on creating textual inversions I did some a couple of months ago, details are hazy, but I'll share what advice I can. Start with this guide: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Textual-Inversion Make sure you're not using any VAE files. Having one of these active will cause your training to degrade over time until it's completely borked. Curate your training set. Try and select each image in your training set to be good examples of what you're trying to train the embedding for. Poor and marginal examples will weaken your results. Check the output of the preprocessing tool. You want to make sure your images aren't being cropped of important detail, or being shrunken to illegibility. Sometimes it's worth resizing images to square shape by hand. I enable the "Read parameters (prompt, etc...) from txt2img tab when making previews" option. This gives you consistent prompt target and lets you judge if the training is going well, or going off the rails. The prompt should embody what you're training the embedding for, but at the same time try and generate something that's not exactly in the training data. This helps you judge if it's more generalized, and not just copying an image from the training set. I like to save images and embeddings every 100 steps instead of the default 500. It makes things very slightly slower, but I like see how it progresses over time. It also lets you better spot when the training "peaks" in its quality. >I've got epic amounts of pics stored I could be getting results out of. Good luck anon.

>>1189 Looks like it might be unswingable for me as I've 'only' got 12Gb of VRAM and some of that is used by Python and background stuff bit of a pig that you essentially have to have a 3090 or higher to do the uber cool shit.

>>773443 >I swear you faggots request the strangest things. It's a /delicious/fag. He has no taste and will happily eat up calarts if someone crudely draws some nipples and a vagina on it.

MusicLM: Generating Music From Text >We introduce MusicLM, a model generating high-fidelity music from text descriptions such as "a calming violin melody backed by a distorted guitar riff". MusicLM casts the process of conditional music generation as a hierarchical sequence-to-sequence modeling task, and it generates music at 24 kHz that remains consistent over several minutes. Our experiments show that MusicLM outperforms previous systems both in audio quality and adherence to the text description. Moreover, we demonstrate that MusicLM can be conditioned on both text and a melody in that it can transform whistled and hummed melodies according to the style described in a text caption. To support future research, we publicly release MusicCaps, a dataset composed of 5.5k music-text pairs, with rich text descriptions provided by human experts. https://google-research.github.io/seanet/musiclm/examples/?utm_source=substack&utm_medium=email Now even the shittiest and laziest of Unity developers can generate unique music for their asset-flip "games" rather than pasting in the Kevin MacLeod library. It won't be long before those flipped assets are also generated on the fly, tricking millions of people into thinking that a game is unique or was made by a human at all. "Automated video game creation" will be a multimillion-dollar market soon.

>>1191 I think there's been an increase in the number of leechers looking for very specific free commissions, and expecting ludicrously complex art done for them that have plagued drawthreads for years now, spamming the same request every thread even after they get what they want.

>>1190 >Looks like it might be unswingable for me as I've 'only' got 12Gb of VRAM and some of that is used by Python and background stuff bit of a pig that you essentially have to have a 3090 or higher to do the uber cool shit. I did some web searching and found there are people saying they've done textual inversion training with only 8GB of VRAM, so 12GB of VRAM should be easy.

>>773526 The perspective in the first pic is all wrong. It makes the bunny girl taller than a house. Her arms are also way too small for the rest of her body. In the second pic, her right hand is stretched and deformed. Her left hand is also deformed despite being mostly covered, what little we see of it doesn't make sense physically. She has a detached tuft of hair floating next to her right hand. Her fishnets are irregular and the shine doesn't match their pattern, so it looks like there is a layer of plastic on top of the fishnet. There is also another weird fishnet on her shoulder which appears to blend into her skin. She has nonsensical artifacts floating next to her head, one seems like a partial white square with a malformed pink heart on the tip. There is also a malformed signature under her right breast, which has the wrong perspective and seems to sag lower than the left breast. Overall, this soulless computer generated garbage has no hope in hell of competing with real artists. The machine is clearly spitting out shit that it doesn't even understand, no matter how much shading it adds to a pic, it has no understanding of what the pic means and which details are important to get right.

>>773570 I'm not the one who is seething. Clearly, you lack the artistic education to understand the kind of thoughtless trash you're posting, filled with disgusting mistakes that not even a bad artist would make. I point this out and you immediately get pissy. It seems to me that you are the one seething because you know that you will never be a real artist.

>>773570 HEY HEY PEOPLE

>this nigger wiped all his posts because some anons critizised him Holy fuck what levels of autism and butthurt.

>>1195 >The perspective in the first pic is all wrong. It makes the bunny girl taller than a house. Bonus.

>>773598 Jesus christ nigger, youre pathetic.

Did he commit sudoku?

>>773605 Yeah, and im calling YOU pathethic, not the other nigger, no wonder you keep coming here to get banned thinking youre some sort of champion of free speech when youre just a butthurt pedophile that wants to post cp on /b/.

At least repost his pictures

>>1203 He didnt take into account anon's autism to preserve every bit of OC.

>>1204 Each time I forget about AI, I see a few random pics and it makes me wanna try it

>>1206 There's a few issues in hands and stuff, but usually 1/20 pictures look very decent.

>>773657 Like i care, AI generated images still look a bit rough but i do love some fluffy/non fluffy tiddies. Plus you can use photoshop or any image editing program to retouch them, so is not a big deal anyways.

>>1209 Anon just be glad he didn't post one with a dick like he usually does

>>773443 >>773526 >>773549 Why did you delete your posts? There's nothing in the logs. >>1204 >>1205 Good man. That giant lava golem is noice.

>>1209 Fox girls can have multiple tails, but that one on the left is clearly not attached to the girl. Not sure what you think is wrong with her wrist either. Why did you wash out the colors?

robots.

i'm having too much fun creating waifus. life's good bros

>>774032 >the male one >1girl your imposed faggotry has no power here.

>>774015 Promotions?

>>1195 Where to begin on your critique... Firstly this is the AI art thread. I expect to see "low effort" content where people just post stuff they've made for simple fun or experimentation, and skipped any serious effort to clean up errors or oddities. >The perspective in the first pic is all wrong. It makes the bunny girl taller than a house. From eyeballing the perspective lines I would say her height comes up to only the edge of the roof. That makes her very tall. Even if you hide the nearby building, she still comes off as a rather tall lass when compared to distant buildings, stones on the waterway's bank, and the flowers. >Her arms are also way too small for the rest of her body. To my eye her arms look the right length for her torso, although they are very slender for her frame. To be honest I've seen lots of artist that draw giant boobs make the girl's lower body seem bigger than their upper body, so they AI could have been mimicking any number of artists when it does that. If there were details I wanted to quibble about in the first pic it's that her cotton tail is off center, there's vertical line where her hair mismatches what's on the left vs what's on the right, her left ear isn't complete, and the art style of the girl doesn't quite gel with that of the environment. Overall I like this piece. It makes me wonder, why is there a mini-giantess bunny girl standing thigh deep in what seems like a small brook? Is it actually that deep? Or is she so heavy she's sunk into the soft soil? How did she get into this situation in the first place? >In the second pic, It looks like initial output of txt2img from the AI. It's like an uncut gemstone, that a craftsman could elect to chip, cut and polish into a thing of beauty, or appreciate for its rough natural form, or move onto another gemstone that is better to craft. So, there are lots of flaws that need to be fixed. >her right hand is stretched and deformed. Yes, her hand is a bit too long. Shaving a bit of the end of her fingers and re-rolling with inpainting should do wonders. >Her left hand is also deformed despite being mostly covered, what little we see of it doesn't make sense physically. Looks like it would be easy to fix with a little bit of inpainting. >She has a detached tuft of hair floating next to her right hand. Trivial to clean up. >Her fishnets are irregular and the shine doesn't match their pattern, so it looks like there is a layer of plastic on top of the fishnet. There is also another weird fishnet on her shoulder which appears to blend into her skin. The AI sucks at doing fishnets with a fine mesh. It does better with a coarse mesh, but I have no luck getting it to do that on demand. My advice to the artist is put fishnets in the negative prompt, and avoid them altogether. >She has nonsensical artifacts floating next to her head, one seems like a partial white square with a malformed pink heart on the tip. >There is also a malformed signature under her right breast All of which is trivial to clean up. > right breast, which has the wrong perspective and seems to sag lower than the left breast. If you assume a point of view that is level with her hips or waist then the sag doesn't seem off. On the whole I think it's a nice image. Most of its problems are easy fixes. However, the fishnets might be too hard to work with, and require starting over on a new image. >Overall, this soulless computer generated garbage has no hope in hell of competing with real artists. Unless there's a collapse in civilization, this technology will just keep getting better, and will "compete" better as time marches on. >The machine is clearly spitting out shit that it doesn't even understand, no matter how much shading it adds to a pic, it has no understanding of what the pic means and which details are important to get right. You're right, it has no genuine understanding. It just mimics things its been trained on, and tries to mimic things that match the prompt, while filling any gaps in ways it "thinks" that will keep the picture "coherent". At the end of the day AI art is just a tool. It's up to the artist using this tool to do clever and amazing things with it.

>>774032 leave this gay and cuck shit, and return to pure waifus

>>774184 Just say you want to post pedo shit and fuck off. Absolute scum.

>>1219 nothing wrong with a cute catgirl friend

>>774623 whats pedo about this? theyre just friends. me and my friends all have double wide hats that we wear together.

making phone wallpapers

>>1222 Looks really comfy, anon! Say, if you're making these, could you make something that's more in line with Future Frenzy from Crash Bandicoot 3? With brighter colors but still a futuristic cyberpunk/bladerunner vibes?

Goddamn I have made so many slutty princesses and aryan harems in a virtual garden it's a wonder my dick is still functional. The photoreal ones are marvellous if you hit the right Steps/CFG, even the uberporn one is really good, though it tries to give tits to every girl when DFCs are required. >>1195 >Overall, this soulless computer generated garbage has no hope in hell of competing with real artists It's made me cum more times than any artist ever has so it's already winning.

What's truly terrifying is this is only gonna get better, my balls are gonna be inside out if they up the resolution and focus on filth. The hottest sloot model couldn't compete even if she tried I can knock up 50 variants of her all coated in jizz in about 2 mins. What a time to be alive.

All I wanted to do was talk about which 2hu I wanted to fug and Remilia had a meltdown. I should've offered to fuck Flan in front of her as proof.

>>1224 >though it tries to give tits to every girl when DFCs are required I suggest using "flat chest" in the positive prompt and "small breasts" and "large breasts" in the negative prompt. That will usually fix it, but for particularly stubborn boobs you can try also adding "medium breasts", "huge breasts" and "gigantic breasts" to your negative prompt. >>1225 >I can knock up 50 variants of her all coated in jizz in about 2 mins. Not sure if you're talking about just making clones, or impregnating said clones too...

>>1227 >particularly stubborn boobs The very worst kind. I'd done most of that but I'd not added thought to add medium and large etc to the neg prompt, I'll give that a whirl, the Protogen ckpt(s) is/are excellent al loli bods though they tend come out a bit high contrast, I need to do a bit more fiddling.

>>1228 Just found this which will probably be helpful to folks looking to craft prompts

>>1224 >it tries to give tits to every girl when DFCs are required. It gives cowtits by default but I haven't had trouble getting small breasts of DFC when I ask for it.

>>1095 Uff

>>776082 sperg harder, r-tard

>>1226 >Obviously doctored footage

Is there any sort option picture extrapolation? As in I input a image like one of the busts in the attached picture and it generates a body for it.

>>1234 Stable Diffusion has outpainting. I haven't used it much because it's tricky to strike the right balance in the parameters between too weak and only getting gibberish, or too strong and ignoring the original image. I've only used it for features like animal ears that originally generated half out of frame. Getting an entire body from a bust would be a fairly extreme case.

>>1234 >Is there any sort option picture extrapolation? As in I input a image like one of the busts in the attached picture and it generates a body for it. The method you're after is outpainting, which is just a fancy way of saying extending an image outward in one or more directions, and using inpaint to fill in the new pixels. The big problem you'll run into is matching the pose with of the face portrait, followed by matching the art style and lighting. I did a quick proof of concept of outpainting a face portrait. Unfortunately, as can be seen in the grid, it had trouble lining up the face with the new image. I took one that was kind of okay, and fed it through img2img upscale and clean it up. The results are alright, but his head is obviously off center. Overall it works, but there's plenty of room for impovement. Prompt and paramaters for outpainting <an old scarred barbarian, digital art, digital painting, witrh a moonlit forest in the background, bald, portrait, full body, leaning, (green sky:0.8) <Negative prompt: lowres, bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text <Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 10, Size: 512x768, Model hash: 579c005f, Batch size: 8, Batch pos: 5, Denoising strength: 1, Mask blur: 4 Prompt and paramaters for img2img upscale: <an old scarred barbarian, digital art, digital painting, witrh a moonlit forest in the background, bald, portrait, full body, leaning, (green sky:0.8), from front, facing viewer, grey hair, (scars:1.2), stubble, claw marks, blue eyes <, (detailed realistic painting:1.0), (shaded:1.3), extreme detail, (detailed eyes) <Negative prompt: lowres, bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, text <Steps: 20, Sampler: Euler, CFG scale: 7, Seed: 6, Size: 1024x1536, Model hash: 579c005f, Batch size: 4, Batch pos: 1, Denoising strength: 0.45, Mask blur: 4

elena from street fighter

>>1237 Too pale, i demand more chocolate!

>>1238 this one came out better i think

>>1239 That is better but holy shit AI really cant do hands.

>>1240 getting the hands right is hard, but you can get the result you want if you the patience to go through enough trial and error

>>1241 >if you the patience if you have the patience*

>>1241 You mean with enough patience I too can post an anime girl with 6 fingers like you?

>>1243 i mean i don't have the patience to try more

>>1230 Think I've got it now, must have been the order of words in the prompts, turns out earlier words have greater weight so moving those to just after the subject got me muh DFCs and fixed a (cum) issue I was having. Good lord I'm creating some filth now.

>>1245 Don't just talk about it, post it.

High-speed coloring workflow using AI image generation [NovelAI] Is in moonrunes, use google or deepl if you can't read it https://note.com/abubu_nounanka/n/ne59633583371 https://archive.ph/W4sDC I also recoment the author twitter for more AI Art stuff https://twitter.com/abubu_newnanka https://twitter.com/abubu_neonanka

you don't see bodies like that in vidya anymore, or anywhere else

>grab hentai >use Stable diffusion to improve it (if necessary) >use AI Interpolation to add frames Would this technically work out?

>>1249 It should work in theory. I don't know if it would look good, but it should work. I'm not aware of any casually obtainable interpolater ais, though I haven't looked. Your main issue with "improving animation" is going to be keeping shading consistent. Assuming you're not just upscaling.

>>1249 SD has no provision for frame-to-frame consistency, so it'd look like a jittery mess.

>>1249 Speaking of this, can you use inpainting to decensor images? I tried out the archived version of DeepCreamPy but it was very underwhelming and apparently no one has bothered to train a newer model ever since the original creator baleeted everything after losing his laptop in 2019.

>>1229 Thank you

>>1252 >Speaking of this, can you use inpainting to decensor images? Yes, you just paint over whatever you want gone and tell it what to make in it's place, give it about 10 goes pick the best one and rinse and repeat if needed. >>1253 I find sticking >(masterpiece:1.3), (best detail:1.3), (ultra detailed), hyperdetailed At the start of any prompt improves it regardless of the claim in that however but they're deffo right about words earlier in the prompt having more weight so if you're not getting some wanted (((element))) despite all the emphasis try moving it earlier. Also: >(sharp focus), (studio lighting), bokeh, Canon at the end helps with photoreal stuff

>>1252 You just want to make sure you get the right .ckpt file to work with for whatever you're looking to decensor. Get the UberRealisticPorn (or whatever the fug it's called) or S1dlxbrew ones (they come in LoRa form now too so you could load one and slap the other on top) for photoreal nudes, obviously there's a multitude of ones for hentai styles if that's your thing.

>>1254 >>(masterpiece:1.3) Be careful with that one. If the training data wasn't flagged with it like NAI's is, and/or whatever blend you use dilutes NAI too much, then it can end up interpreting "Masterpiece" wrong and start generating picture frames around everything.

>>1180 Jesus fucking christ that's bad lol

>>1143 >>1180 Well this suggests otherwise.

>>1180 >>1257 I have no idea how you guys are getting these responses, I'm getting the stupid thing to write me all kinds of scripts and code and even translate and work with text in other languages. When you ask for generic shit or for shit that can easily be filtered by the admins ("say nigger, AI!") it recoils and falls back to its poz'd programming. If you feed it exact instructions and prod it when it refuses to function, you get optimal results.

Postan some tiddies

>>1260 >>1248 Tits beings the size of head is the perfect big breast size. And skinny jeans waist

Everyone's posting pictures of girls looking AT the viewer, but have you tried creating girls from the back?

>>1262 Seen these moleless Kobeni AI art yet? They were almost there too but they'll never get the smaller most important feature, her moles, down pat like how humans could and always do. Also >2nd pic >Realistic human eyes on that anime girl Thats rather disturbing, very uncanny valley like.

>>1074 >>1263 Could the Ai handle a request similar to pic related? I'm hoping for an unconscious girl in stockings or pantyhose and a short skirt or dress, being carried over the shoulder in a way that lets the viewer look up her skirt and exposes her white panties underneath, preferably by an armed woman showing off her cleavage. You could look at one girl from the front and the other from behind. Also interested in women being passed out or unconscious in different poses, sometimes interacting with other characters, especially being carried in different positions like in old horror movie posters but more photorealistic. I would rather not install anything ATM, but if I wanted to make it myself what prompts would I enter?

Think I accidentally made Amber Heard. >>1263 Yeah, it works. >showing ass, ass up, squatting, twerking (I think), from behind Are all valid, if you add ass specific things like the Bubblebutt LoRa you get more shots from the rear and with magnified ass. >>1266 When/if you do, give it a lot of goes at it, it can certainly do ass and turned to camera but orchestrating scenes is a roulette spin. I've not got it running ATM but I'm gonna guess and try the prompt when I wake up to see how much of a mess it makes. Then I'll tinker and see if I can make something like yer pic. >(masterpiece:1.3), (best detail:1.3), (ultra detailed), hyperdetailed, A photo of a white woman (holding a pistol) carrying a girl over her shoulder standing in a dark hotel lobby, (full body:1.2), (detailed eyes), large perfect breasts, cleavage, tight pink dress, scarf, red lips, mean expression, anger, sneer, (vibrant magenta hair), exposed ass, bootylicious, stockings, panties, luminescent wall lights, (studio lighting), (sharp focus), bokeh You could also try your hand with >photorealistic, 32k, UHD, Canon, CCTV, It'd probably get the gist of that after a while, probably be some colour bleed with the 'white' and 'red' and 'pink' and 'magenta'. the Neg prompt would be a suite of standard ones about shit hands, extra limbs and poor quality plus anything undesirable that crops up over and over in the process of churning them out.

Hilariously if you put nudes in a garden they come out white or asian, put them in a park and they become latinas or hues, put them in prison and yes they're nogs.

>ai can generate pregnant belly >but thinks tan is a shirt

>>1269 Just put shirt or clothes into the negative prompt.

>>1268 that's hilarious, post some

>>1263 >>1267 How well does it handle women's underwear and genital areas? Like let's say you asked for a photo of a girl bending over in a leotard or miniskirt, shot from behind, with a closeup of her pubic area. I know it gets clothing including underwear right for the most part. Could you also ask to see an outline of her labia underneath and specify how much is actually exposed or just outlined?

>>1272 Now there are so many different models, loras and what not that you can do anything .You can use inpaint to have the ai edit only a certain part of an image, the possibilities are endless

Have you ordered your GPU yet anon? You're not going to wait until you need to get in line to suck Biden cock and get your gaymen GPU gov permit, right?

>>1274 Yeah I'm still not spending $9,000 on a GPU. I like AI but not that much.

>>1272 As the other anon says: It 'can' do genitalia well, it can just take a lot of goes to get it right, *if* you want to do a significant amount of InPaint/editing job to get it, you'd be very lucky to have it generate something that specific without meddling. Genitals are a roulette spin, sometimes they come out looking like there's a knuckle between their legs where they've been trained on shots where girls are masturbating

Mostly I was hoping for images similar to pics related with detailed photographs of women lying unconscious in different poses, only more risque with things like exposed underwear and sub/dom overtones like >>1266. I'm mostly concerned with genitalia so I can control the way they look under the underwear for upskirt shots and similar. Also hoping to be able to control the girl's expressions since I really like the closed/eyes open mouth combo and also the dangling legs and exposed necks. I know scenes are a roulette spin but how would I have two different-looking characters interacting, like in the first pic? I'm guessing it would be easier to generate the two images separately and then have the AI combine them. Would it do better if the characters were similar looking, like in a picture with two women lying down in the woods together, or one woman standing over another? I don't want it to start applying the same set of prompts to every character for obvious reasons.

>>1277 Not an ideal checkpoint for realism but it was a bit more malleable with the colours, couldn't get her carrying an exposed ass but it's very apt to subvert/reinterpret a prompt weirdly. It certainly prefers generating a primary character, or multiples of that character but can make 6 to 8 distinct people in a group, more than that they tend to get a lot of similar features or outright sameface and/or blurry. Expressions are controllable yeah it's mostly down to ((emphasis)) or using synonyms to get some of them working. As you say it's probably best to Inpaint to create scenes where you want specific interaction between characters or else leave it running for ages to see a selection of shots you can play around with. Combining two distinct images is something I've only had limited success with, the results tend to be over-stylised to hide the differences between the original shots, that said I I've not tried de-stylising the results with a photoreal checkpoint yet so that might work.

>>1278 What about an image of an unconscious woman with no other characters? Like let's say a portrait of a woman with reddish-brown hair wearing a wedding dress and bridal veil and lying down outside on a forest floor.

>>1279 That's a complicated scenario anon. How many times have you fantasized about it?

>>1280 I was hoping for something like picrel as a photo, only outdoors. I thought that a portrait of a woman who had fainted on her wedding day would be simple and fit what I had in mind. I imagine the photo that had taken just after she had collapsed but before she had regained consciousness. It's not a scenario I've put a lot of thought into, she could just as easily be a bridesmaid, a nurse, a secretary, a schoolgirl, a nun, a fantasy princess or pretty much any type of female character that you could imagine in a similar situation.

>>1065 > The amount of People who glow in that website is unbelievable Hey, even the glowies need a place to hang out and shitpost. How would you like it if half your job was procuring middle aged girls for Joe?

>>1281 >picrel

>>1281 >pic nice

>>784427 Such a shame you'll never bury your face against a warm prime cunny, instead you'll resort to wasting away your life generating plasticine shallow cookie cutter child models. I bet you don't even generate actual cp you fucking low-test coward, that lack of testosterone is why you're a pedonigger in the first place, because you know you can't keep up with a real woman.

It can be really fun to play with

I don't have enough autism to make it look that great

>>1286 >>1287 What were the prompts for those images if you don't mind sharing?

>>1287 Looks quite good to me

>>1289 Oh, except whatever going on with the fingers I guess

>>1288 > Detailed Lines, Golden dress, Odin, ((Skindentation)), ((thick_thighs)), ((large breast)), Multiple realistic Wings, blue eyes, Detailed Face, cleavage, Feminine, Short Hair, Solo, highly detailed pupils, detailed lines, white hair, sad eyes

SD can produce such weird and wonderfully degenerate things, and all people do is waste their time making generic anime girls that are virtually identical to what you can find on any random Pixiv profile. I've seen the LOL threads, I know you lot aren't this boring. Show me what you've got.

>>1293 I wish there was a better version of the second pic. Seems like a good base.

Has anyone tried generating photorealistic female faces? Anime girls are all fine and well, but what I'd love to see are some wallpaper sized portraits or profiles of beautiful women in different situations and settings. Not necessarily anything overtly sexy, though cleavage and pantyshots would be welcome especially if they make sense in the context of the image. The more variation in clothing, posture, emotional state, ect. the better.

>>1295 I was playing with that too

>>1297 Nice! Looks looks some sort of cyborg girl?

>>1299 2, 3 and 5 are awfully cute.

>>1300 Agreed, protogen really seems to do cleavage well. The only thing missing is more panty shots.

And that's enough for now, hoping to see more of these images in the future.

>>1303 So, for the pictures which obviously look like certain celebrities, was that intentional, an artifact of training, or just the AI trying to fuck you?

>>1303 >infinite pictures of prime age Christina Hendricks fucking hell my cock could not be harder