Previously, on /asdg/ >>8561383The cartoony side: >>>/aco/csdgNo DALL-E, hailuo AI generated stuff.No futa.>>>/aco/futasdgReport trolls.>GuidesGeneral: github.com/AUTOMATIC1111/stable-diffusion-webui/wiki | rentry.org/sdg-link | rentry.org/rentrysdHires Fix: rentry.org/hiresfixjan23Inpainting/Outpainting: rentry.org/drfar | rentry.org/inpainting-guide-SDAnimation: rentry.org/AnimAnonCreate LoRAs, TIs: rentry.org/59xed3 | rentry.org/simplified-embed-trainingList of models/LoRAs:rentry.org/acorepo>Auto1111 extensionsControlNet: github.com/Mikubill/sd-webui-controlnet | rentry.org/dummycontrolnetRegional Prompter: github.com/hako-mikan/sd-webui-regional-prompterTag autocomplete: github.com/DominikDoom/a1111-sd-webui-tagcomplete>Related>>>/b/degen>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/g/sdg>>>/trash/sdg>>>/trash/slop>>>/vg/aids>>>/vt/vtai>>>/y/ydg
And that was the usual archive dump.
i will try to edit her tit>>8571414stunning
Last repost: does anyone have, or know where to find, this mysterious lora:>crop_shirts: cd390067ede3I'm trying to reproduce an image that references it, but I can't find it anywhere.Googling gets me only two japanese hits with the same in the metadataI thought I had found it on tensor but didn't notice the hashes didn't match and it fried the image
Queen Amidala Dreaming.
more big tit abby, less shitted
>>8572103nonsense, abby was made to be blacked, all jewish women are naturally attracted to blacks
>>8572172okay but that's for IRL and this thread is for porn people actually want fap to
If requests are cool, can anyone do Allyson Hannigan circa Buffy in the 90's, and/or Kiernan Shipka?
>>8572113This but hairy pits and forearms please.
>>8572239might be able to on thursday or friday
>>8572373Sounds good, thanks
>>8569701Phenomenal.
can sorta do gangbang, but this model and lora is tricky for me
>>8572722meant for >>8570825
>>8572723
>>8572722>>8572723>>8572726Can you do ScarJo or Emilia Clarke like this?
>>8572823>>8572824Nice.
>>8572722>>8572723>>8572726something about this jew, dont know what is it
>>8572827the cat is killing me
>>8572833
>>8572840
>>8572944
>not a single cartoon genThat's not very /aco/ anon.
>>8572748def emilia clarke. I don't have a scarjo lora for biglust yet
>>8572998You asked for it.
>>8571400Please more just like this! Exposed, wearing helmets, huge breasts, and bush!
Hi /aco/ I was told that anons here have knowledge on training loras of real people, sorry for imposing.I've been trying to create a lora of a real person since SD 1.5 (currently attempting with realistic Pony mixes) and no matter how I do it it never goes well, the body gets learnt perfectly but the face is always off, always feel like either undert rained and over trained.I've attempted to train with cropped images where only the person is there, full images, booru like captions, sentencing captions, and I don't know how many combinations of models , schedulers optimizers and network dimensions and alphas variations.I always see the epoc/loss graph decreasing meaning the learning gets done but... hoinestly I'm kind of lost at this point.I'm kind of limited because I have only 8gb of vram, but I guess if the training starts I should be able to find the right setup for it to be good.Anyone has any advice here?
>>8572998Newfag or troll>>8573080The loss graphs can be highly deceptive, it's not an exact scienceAlways fix the training seed, and sometimes you can just reroll a different seed and get a better LoRA from the same training set and settingsThe quality of a LoRA is also deceptive sometimesAlways use with the checkpoint you trained on (the common wisdom of using the base model is bullshit if you want good likeness without stiffness)When applying the LoRA, you can set TE and UNet weights separatelyPlay with them separately to figure out which is over or undertrained and adjust learning rates accordinglyTry tweaking your training set: fewer images is often betterStick to a similar look in terms of age and makeup especially
>>8573193imo the training data is most important. I've trained thousands of models, SD and orherwise. The thing that impacts training the most is your dataset. It's probably 80% of the work. The last 20% should be tweaking the hyper parameters.
>>8573193>Newfag or trollThat's rich, you must be too new to know how it used to be like if you think it has always been like this.
>>8573225This was always a /tv/ refugee thread, never much for comic stuffIf anything, there should be more celeb gens againCartoon characters simply don't look good in realistic form
>>8573223Remove tattoos, give pussy hair.
>>8573249there probably could be a celeb general but that shit is about be illegal so im guessing thats why it isnt posted all that often
>>8573274lol
>>8573278i know right, that would violate free speech. What is anon even talking about. Go back to posting Abby bbc paizuri
>>8573249back to your failing dalle threads so you can spam 60 of the same fully clothed boring ass celeb with you. good brainwashed simpleton. worship thy false idols. vote as they order you.
>>8573080
>>8573340Heh, thanks, I love it!>>8573193Thanks a lot for the help.>fix the seedNever ocurred to me, will try>less images my biggest set now has 235 images, I'll try reducing the scope upping the repeats>TE and UNet weights separatelyNever considered that approach, it seems I still have a lot to learn, wil do so.In terms of epochs / steps, I've seen a lot of variance in the recommendations from different guides, faqs and posts, from jsut 5 -10 epoch to up to 200 and pick based on the epoch / loss graph, any insight on that aspect?Somethign that has been happening lately is that I train for 10 epochs / ~11k steps and the output of the LoRA is the same on the fisrt save than in the final one, I've attributed this to having a network rank too low (from 64 and below it seems to happen), but only when I train with the Prodigy or Adafacto optimizers, never happened with AdamW8 or AdamW.
do you manage to do perfectly formfitting clothes / suits AKA "impossible suit".I manage to do coating, layer, second-skin, but it tends to apply to the whole body.
>>8573380skinsuit?
>>8573385skinsuit is okish, but it tends to constrain curves, and also how would you make it to a swuimsuit / leotard...I'd like this type of fitting
>>8573385>>8573395Also I want completly uniform and seamless, which is why I tend to go with coating/layer/secondskin
>>8573395Maybe "latext body paint" or some weird shit like that will get you that effect?
>>8573414yeah, coating / layer, works, but I can't restrain it to a swimsuit shape. Also latex tend to generate seams
>>8573421so just like a smooth casting, I see what you mean, and I guess thre is no Lora for anything like it I presume.What about inpainting?
>>8573299You realize comic characters are owned by corporations too, right?I can coom to pictures of celebs, I don't listen to them...Degrading them in self made images is the opposite of worshipAlso you shit on dalle while posting the most dalle-looking gen ever, bravo>>8573360I train with 6-30 images for 2000 steps max (split into 8 epochs), rank 48 is enough for meSometimes epoch 5 is already trained enough and so I end up using that
>>8573544
are male gens allowed here?
>>8573682or at least, which board has the male equivalent of this thread of photorealistic AI men?
>>8573683there's a couple of homo boards
Ok boys, does Sd 3.5 know celebs? I need to know if I'm wasting my time or not.
>>8573772give a little more time so the people that do make those loras and stuff to make them. 3.5 JUST released
anyone know where/how I could find high res ms rachel photos for a lora?
>>8573035
>>8573541Any chance for a catbox or used checkpoint, loras, and prompt?
>>8573457>6-30 images for 2000 steps max (split into 8 epochs)Then I guess I'm overtraining by a large margin.
Some snu snu
>>8574056Maybe the face is not consistent enough between all 235Keep in mind that the training process works like a giant Venn diagram: it teaches the model all the things that are common among all the imagesTry starting small, with fewer images (maybe face only)Overwrite the tag of an existing celeb that looks similar so you get a consistent body (and faster training)
>>8574071
>>8572748eh
>>8574141
>>8574076WIll try, foir now I've divided my set into >face>body_dressed>body_cosplay>body_naked>sexDid a bit of further selection adn reduced the amount of images, provided different repeats based on the number fo images of each group, testing the train agian.In addition, which shceduler and optimizer do you recommend? the thwo most common combos I see are cosie+prodigy or constant + adamw8bit, for those people that do not use custom schedulers that is.
>>8574177>>8574076Sorry for the inquiries, but I seldom find someone that knows about these things that is willing tio share / help, do you use regularization images?
>>8574142oh no whos that on the back
>>8574177>>8574188Just use constant schedule, anything else is snake oilIf the learning rate is too high, just lower it and train longerThere is no point in slowing it down during trainingI just use adamw8bitRegularization is not needed for this, there is no gainYou are not training any new concept here, it knows full well what a woman isYou are just tweaking the details a little bit
Please someone tell me the prompts please, I'm trying but it's not coming as good as guys are, wanted make a image where a big tiddy girl opening her legs and thick cum dripping out of her with a grin smilehelp me
>>8574203if you dont know how it works, just put your phrase into AfroditeFlux
>>8574203This isn't Dall-EIt all depends on checkpoint models and LoRAsCheck the metadata on a civitai image or ask someone to upload a png to catboxIn A1111 you can import all settings from an image with the PNG Info tab
>>8574201reduced my set to 114 images, 5 repeats for each set.>25 for body>30 dressed>30 cosplay>29 for facemight still be a bit too much? Another thing I usually need to cache the textencoder outputs because my vram fills if not don't know if that might be the cause, but not all the time and I don't know which config changes this.
>>8574205dude you are Godly!! now can you make this looks like a black cock cover here eyes, her mouth is open with tongue saliva dropping from it, now the legs wide open like a siting and the cum should be more thick and wide and it's oozing out of her like a cum dumpster
>>8574206can you teach me how to do it like I am 8, it's very new for me so
>>8574207Looks like a good amountDunno about those settingsI just stick to 1.5 training due to VRAM limits and haven't tried harder to get it to work
>>8574209Man, this is not /r/.Just read our responses
>>8574213okay then teach me
>>8574200Don’t recognize him without his funny little hat… post more though
>>8574210>>8574215Nobody is gonna walk you through every detailMake gens, gens, gensImprove and learn over timeIn the beginning everybody wants to make very specific stuffJust try to make good looking stuff firstThe AI can sometimes show you things you didn't know you want
>>8574225not very helpful I guess
>>8574236Just gid gud, scrubAfter a couple thousand gens, things will start to make senseThere is no shortcut, AI is dumb as fuck
>>8574203i'd recommend looking at prompts on civitai, and mix and matching them with what you're trying to accomplish. ponyrealism might be a decent checkpoint to use since i believe it uses booru tags.>>8574200i need a blender model of her...
>>8574240can I use civitai for all adult purposes?
>>8574245No celebs, no scat and another thing I don´t remember ATM.
>>8574245civitai is a website, and yes. but you need an account to post and access lewds i believe.
>>8574246>>8574247it asking me to sign in with my discord and also how much money you guys spend a month to make your desires pics?
>>8574248it also has google/email sign in's. also i don't really use that for making renders, most of us here use comfyui/webui so it's free. if you haven't even got that set up, i'd recommend searching up yt videos on installing it.but civitai can be used to test making renders.
>>8574248>>8574248dont put money on website based rendersDownload and setup AUTOMATIC1111 on your pc, and get models on civitai
>>8574252>>8574253Thanks, give me sometime mate, I will do all of those and tell you in a couple of hours did it work or not.also do you guys know something same as this where I can generate videos with no cost?
>>8574211Ah! 1.5, I'm trying for SDXL which gets tricky with 8gb, I'm now testing a new training wiht fp8 see how it fares.
>>8574255>>8574255This is all novel tech, we are all fighting our own battles. That said, I just got set up again after a few-years hiatus, here's what I did to get pic related:>download AUTOMATIC1111>read error messages, address themNote: I'm on an older Nvidia card with 6GB ram, so the following worked for me but won't for you if your setup is different:>in webui-user.bat (READ THE DOCUMENTATION, GOOGLE FOR STUFF YOU DON"T KNOW), I set set COMMANDLINE_ARGS=--skip-torch-cuda-test --lowvram --no-half --xformers --no-half-vae>used Euler a (is this still the best?) and biglust v1.5 (someone here mentioned it)>used X/Y/Z plot script (bottom of page) to vary CFG scale and denoising strength, testing out possibilities, to find what works with my proomptAnd here's the proompt:>positive: athletic, tall, blindfolded, 18yo, perfect, beautiful, [curvy], fit, abs, full body, front view, 1girl, solo, nude, naked, [perky], nipples, pussy, thighs, boudoir, realistic, sweating, natural, hires>negative: low quality, sports, advertisement, fat, overweight, obese, bbw, anorexia, implantsIf anyone who hasn't been living under a rock the last 2 years has any tips for improovements, I'm all ears. Right now I'm testing steps and hires steps, next I'll try adding a ballgag to make the mouth less weird. Iterate, iterate, iterate, iterate, iterate. When I was set up well my x/y/z plots would iterate over celebs, poses, and outfits, because everything else was all set up, it was cash. Not looking forward to the power bill. If I could ask someone here for anything, besides general tips, it would be how to cover the face entirely with a blindfold so I don't have to optimize over it, my bodies are basically perfect. Or how to do faces, all your faces look amazing so maybe I missed something big.
>>8574305For normal gens in SDXL I find --opt-sdp-no-mem-attention is working better than --xformers. It's faster and resolves quicker. I have 12gb vram though.For specifically SD 1.5 upscaling xformers is still faster.I think most people use DPM++ 2M for the sampler now.BigLust is great and a real gamechanger for realism. Also check out CyberRealisticXL.
>>8574305change o forge instead of AUto1111, you an use SDXL models even with 6gb of memory, I dop on my 1060 6gb while I attempt to train loras on the 3060.
>>8574378>>8574314Thanks bros, I'll give it a try. I don't even know if I'm running SDXL but holy crap biglust is nice. DPM++2M gives similar results but faster, so thanks for that as well. Do others find increasing steps, both for initial and for hires, helps a lot? Maybe I was scraping too low but damn this shit is excellent (picrel).
>>8574393Higher steps do help, up to a point. Check your model/checkpoint to see if it has a recommended range. Advice also valid for CFG.
>>8574093amazing! if i close my eyes i can fap to it!
>>8574393On civitai and hugginface tghe big / good models usually have some instructions / recommended settings for samplers, cfg, clip skip, and prompt advice if needed.Also check compatibility of loras with the models they have been trained on, they suualy work well with those and mixes of those models.
>>8574378Does Forge do everything that A1111 does but more and easier? I'm mainly thinking about the ability to directly inpaint within the UI which ComfyUI didn't allow.
>>8574507yeah better and easier it also has more development, auto1111 has slower maitenance commits.
>>8574314did you make a sweeny biglust lora? can you share?
>>8574524Look for the 217.87MB Sydney Sweeney XL Lora on CivitAI by rumpelstiltskin for SDXL 1.0
>>8572239>>8572239i have something rolling here
>>8574545Hairier please.
>>8574305Faces are fixed up by an extension called adetailer, which fully automates inpaintingYou can stick to lower hires steps (10-15?) which speeds up things quiet a lot without affecting the qualityThe DPM++ samplers are standard really, with Karras noise schedule3M takes the biggest internal steps and so might need a lower CFG to prevent the image from looking cookedDynamic prompts and wildcards are the standard for mixing up promptsEmbeddings have fallen out of favor and LoRAs are the standard nowUse that for accurate celebs, good ballgags etc.
>>8574305>>8574600The sampler algorithm depends a little bit on the used model. Some models work better with a simple Euler Normal, others require DPM++ 2M SDE Karras. Play around a little bit and check the sites where you get the models. Usually there is a short description telling you which algorithm should be used with the model.
>>8574621Euler is sort of the base line, but it's slowThe DPM samplers take shortcuts and so they run faster while still producing very good outputPeople read way too much into the sampler choiceJust read up on the technical side
>>8574545>>8574550>>8574551Damn these are amazing! Thank you! Is doing Dawn Summers from the show possible too? Either way thanks again
>>8574658There is no SDXL Lora model so Michelle Trachtenberg cannot be BigLusted!
>>8574658>>8574672you have any specific requests for hannigan? Could do a few more. I can do about any celeb, it just takes time to train a model because I have to collect data, edit the images, caption them, run the training etc. Side note I might do a Michelle Trachtenberg 2004 lora
>>8574676I've never trained a Lora. Are there minimum requirements and is it easy?
>>8574676You're knocking it out of the park! Anything with her fingering herself, giving a handjob, or a blowjob would be hot as hell.
How I can create my 2d picture in motion just like in immersity ai, make it lively?
>>8574679I'd say in general it would be good to have a newish nvidia graphics card. You can do it with as little as 8gb, maybe a bit less, but it will take forever. 8gb is the minimum i'd say. You can also run trainings on civitai and other providers. As far as easy - I'd say it's easy if you have someone else's training parameters to start wtih. If you're starting from scratch running this stuff locally, and you aren't very savvy with computers and software, then it will be tough.
>>8574676Ooh. I wanna see both hannigan and trachtenberg as cheerleaders sucking lots of cock!
>>8573360If you can, post a dataset and I will try and bake a lora - I typically do a handful per day, more than actual genning.
>>8574743I really appreciate the offer, but the only dataset that I'm working with are private pphotos of my wife so... I hope you understand that I can't share them.Do you bake for SDXL on 8gb?
>>8574747I have 16gb but I gen while I bake so I have the model loaded 2x in Vram :) I can bake using only 8bg - I usually run batch size 4 for speed and could lower that too.
>>8574770Yeah but with 16gb it won't break the training if it requests 20mb more when its at 7.9gb capped.I gen with my 1060 6gb while I train, it's a pity I cannot use boith to train, oh well... I'm waiting for the 5 seriues to see if to pick a 5080 or a 5060Ti 16gb (if they do teh same as the 4 series).
>>8574775My driver swaps to Ram for a few seconds when the VAE kicks in, but it only slows the training for a bit, doesn't break it
>>8574780huh, how? with Kohya or onetrainer?Foir generating Forge uses ram on the 1060 sometimes and it slows to a crawl but it finishes the gen, how can I achieve that for training?
>>8574786I think it's a setting in the Nvidia drivers, I toggled it once and it Webui or Kohya crashed when they ran out of Vram, now they just slow down.I always have performance/Vram tab open in task manager when I gen & bake.
>>8574794Forgot to add, my training usually breaks before it starts, after loading the captions if I don't cache them or right after it does 100-200 steps suddenly requests more mem and stops. Oh and alñso if I set up the dim to be over 64gb, it will break as well right after the captions load.I need to find some weird configs and so far I have never succeeded in an output that it's to my liking or close to the good ones I see online.Most of the time it seems that the network is not learning anything though.>nvidia drivers settinglooking for it right now.>I always have performance/Vram tab open in task manager when I gen & bake.Yeah that's a given.
>>8574795I run Dim/Alpha 4 for characters and concepts and 8/8 for styles
SD3.5 can get somewhat nippy with no loras.
Oops. Better image.
>>8574796for SDXL? such low dim? wow I have to test that then.
>>8574796
>>8574802Yea I mean I'm *compressing* like 3mb worth of jpegs into latent space and it takes 217MB? hell no, someone didn't do their homework.SD1.5 basically knows *every* place and person and artist and car type and movie and animale and plant that ever appeared on the internet and was 2GB total
>>8574771wow an even jewier allison hannigan
>>8574833her name is 1girl
When using adetailer for faces with hires.fix, it tends to change the face a lot compared to the non-upscaled versionIs there any fix for that? It's kind of frustrating when you have a nice facial expression in the base genThat's with ESRGAN_4x, 16 hires steps, denoising 0.1-0.4
>>8574881In hires fix use the Lanczos upscaler (just not Latent) and change the denoise to 0.4 or less then use around 0.3 for the ADetailer inpaint denoise strength.
>>8574893My adetailer denoise strength is usually 0.45-0.55Going lower as you suggest might do the trick, thanks
>>8574881With Pony this is a real problem and I only recently found a fix. Notice how Pony faces get horselike (duh lol) when they fill the frame and more normal when they are smaller? Adetailer always fills the frame so you get teh distort.Solution is mask dilation 80, padding 80. I use denoise 0.4, only higher if I need to change the face a lot (face swap). Hires upscaler doesn't matter much if you denoise at 0.6 anyway, try none and your fave and try to see a difference
>>8574551neat grainy/filtered looklike a 90s magazine
>>8574771Nice! Thanks for fulfilling the request
>>8572827Gotta love those steampunk tiddies...and kitties.
>>8574774what models did you use for this? so good
>>8574954its biglust 1.5 and a custom lora
>>8574924what model is that?
>>8574817me on the left
>>8571396After a while, all american women in movies and tv shows start to look the same.Maybe everyone on tv or disney movies in the last decade are simply simulations of living things.When I browse CIVITAI all the women and LoRAs and actresses and porn stars look the same.Hermionie Taylor Johanssen... are they even real. Why should we care if pixels get raped or shot by made-up stalkers?They're stored on hard-drives in 8K.
>>8575020I have lost faith in tv and movies.Maybe even the news isn't real.Does the usa even exist?Maybe it's just Hungary + CGI. You never hear any news from Hungarians and nothing seems to ever happen there... maybe they're busy pretending to be the usa so hard they're too busy to have natural disasters and murderers and bank robberies?I don't think the usa even exists. Maybe it was a joke on a Polish late-night talkshow that got way out of hand.I've never met anyone from Nebraska.... maybe it's made up... like Nevada, Narnia and Wakanda.
>>8575028Pretending the usa is real is a mental-illness.This house isn't real.... just like florida.
>>8575037enjoy your shirah law.
>>8575079
>>8574747>>I really appreciate the offer, but the only dataset that I'm working with are private pphotos of my wife solol why do you want to make virtual porn of your own wife?
>>8575336Nice hair, isn't that an already existing porn pic though?
>>8574822Yeah, it makes senseI found out the setting in the nvidia driver, but it¡s not helping>Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward passtorch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 10.00 MiB. GPU 0 has a total capacty of 8.00 GiB of which 0 bytes is free. Of the allocated memory 7.13 GiB is allocated by PyTorch, and 166.96 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONFI guess I'll need to postpone training until I get a card with more mem.
>>8574799>>8574801cute faces. are these based on real people?
BigASP 2 got released today.https://civitai.com/models/502468?modelVersionId=991916
>>8575564Just stick to 1.5 for nowHarder to gen good stuff, but faster to gen and train LoRAs
>>8574924This one was CyberRealisticXL.
>>8575706nice pussy
>>8575711my favorite of the set
>>8574991>>8575682oops wrong quote>>8575663tried it but Big Lust seems to produce better results without all that autistic prompting in the Big Asp description.
Are we allowed to post scat gens here?
>>8575336Not that I don't appreciate this trio of photos, but any chance you can re-upload them but with color?
>>8575079almost looks more like an Italian ex of mine than Abby... damn... No wonder I find her so hot.
MOAR ABBYYYYYY>PS: >>8575752 disgusting, never gen again
bigasp is quite sensitive lol
TIL Big Lust seems to accept score tags. Not a fan of that system, just create good results all the time?
>>8575770i could probably colorize them somehow. They were genned in monochrome/grayscale
Does anyone know how to make the A1111 Style text appear at the start of the prompt? Big Lust and related models seem to benefit with score tags at the start of the prompt but a Style gets appended to the end. Text sequencing can make a large difference.
>>8575770>>8575783ControlNet supports recolouring.
>>8575800nice that is dope - is there anything special to use that controlnet or is it just a matter of load the recolor model and use it?
>>8575781How the FUCK do you get faces so good. Adetailer is a coinflip whether it's any better than without it. Do you just reroll until you get a good one? Using mediapipe_face_full if that matters. Right now I'm testing full face masks, hoods, blindfolds, etc. so I don't have to deal with this, but if I can get semi-reliable (I know nothing is reliable) faces like these that would help.
>>8575777You know what ? Im gonna gen even harder
>>8575808ControlNet extension enabled with Recolor selected and ioclab_sd15_recolor.safetensors placed in the model subdirectory within the extension folder.Can generally leave default settings and the colour strength can be controlled by the denoise strength. 0.85 was used above.I also used the CyberRealistic v60 SD1.5 checkpoint so I guess something similar would be fine.
>>8575793Use the xyz plot script with prompt s/r, put the style prompt in parenthesis so it's treated as one prompt and put the substituted word at the start of your prompt. So "blstyle, 1girl, solo, oppai loli" will be changed to "(score_9, score_8_up:1), 1girl, solo, oppai loli" if you put blstyle in the x values field and select prompt s/r under x type. Haven't tried it but that's what I would do.
>>8575839Shit, meant to say blstyle in the prompt and the style prompts in the x values field.
>>8575830I'm experimenting with score tags for Big Lust (picrel). I think the one above was just hires fix without adetailer. Face correctness can be a bit random so either just adetailer, just hires fix or both like picrel (she still got a wonky nose and a mark which may or may not be a tattoo). ~30 steps also helps when the face is further away. I'm also playing around with DPM++ 2M SDE with a 4 config scale.
>>8575682Thanks anon, I like it.
>>8575881
Is there a "Civitai" for XXX celebs?
>>8575941For generation? No AFAIK, and pretty much unlikely to survive if it does. Collecting images? Also no AFAIK. You have to check everywhere from rule34 (dot) xxx with the tags ai_generated and celebrity to celebjihad.
>>8575916>>8575881>>8575906
>>8575793Use {prompt} in the styleIt gets replaced with the prompt, so you can append and prepend stuff in a style>>8575830What model are you using? Maybe it's just shit at facesAlways "Inpaint only masked" checked and don't crank the padding up too highDepending on model, sometimes I used a separate prompt for the face that includes something like "closeup,professional photo,modelling headshot"
>>8575978Perfect, exactly what I needed, thanks anon!
shalom
>>8571403oh my me likey her naked
>>8574886finally an adult cartoon thread I can get behind
>>8575800Can you do a hairier Elaine please?
>>8576357Forget King Of Camelot, this is Queen Of Cameltoe!
>>8576357Cheese
>>8576355she can touch my prostate with half a finger
Full disclosure, I am retarded.Trying to gen using a refiner. Getting an error requiring me to use --no-half argument in Automatic1111. So, I instead launch with the webui-user.bat and get all sorts of memory usage errors including a pytorch cache error. Now, even if I remove --no-half from the startup arguments, using webui-user.bat will crash me out.I can only use webui.bat now and it seems like the threshold for a memory error is much lower now. What gives?
>>8576392What gpu anon?Also, maybe try with comfy?-> https://www.reddit.com/r/StableDiffusion/comments/1gb07vj/how_to_run_mochi_1_on_a_single_24gb_vram_card/
>>8576235>>8576251>>8576308i'm tired with her. show us other celebs
I like the lighting here
>>8576396RTX3080 with 12gb VRAM. I tried clearing pytorch's cache and that didn't seem to help
>>8576404Bake your own shit faggot
>>857641212GB isn't enough for this unless you with 2 seconds clip or something.Look here : https://github.com/kijai/ComfyUI-MochiWrapper/issues/2
"I will fight no matter what week it is"
>>8576392Sounds like my experience in the beginning, running out of VRAM sometimes and things that used to work stopped workingWhat happened was that I tried to do something that my card barely had enough VRAM forThe OS also needs some VRAM, and so it just so happened to work once or twice by pure chanceBut bottom line was that I didn't fulfill hardware requirements (SDXL hires upscale >1.25 in this case)
>>8576417I'm talking about txt2img, not img2vid or txt2vid.Issue arises when trying to use the refiner to take something from anime/cartoon to photorealism. No issues doing this utilizing Fooocus or Ruined Fooocus. Doesn't play well with Automatic1111. I'm wondering if there's a better method for turning anime into photorealism?
>>8576516>I'm talking about txt2img, not img2vid or txt2vid.Oh sorry, yeah in that case it's unrelated.
>>8576516img2img my man (maybe with controlnet)You are misusing the refiner feature here
Wanna play a game, Anon ?
>>8576583Do you have any suggestions? I've not had great controlnet results but might be a skill issue
>>8576604canny and openpose should workMake sure you have SDXL controlnet modelsUsually when nothing is mentioned in the filename, it's a 1.5 modelThough the XL ones are quiet memory hungryBut depending on the cartoon/anime, even regular img2img might get you far if you tweak the denoising strength and have a good prompt to help the model outRemember you can replace the base image with a generated one and run img2img repeatedlyI wouldn't expect a perfect image from a single shot gen
>>8575906>>8575916prompt and Loras used? This shit is top tier
>>8576416keep your ugly jew for your little dick
>>8576404fuck off retard, Ive been waiting for years to see this whore getting dicked and this is the closest we will get
>>8576404I'm not, more Abby!
>>8576679>>8576704based .
>>8576656
>>8576679>>8576704
>>8576907Niice
3.5 is pretty cool
>>8576584What board game is that? Or does AI random image generator skew the image of some obvious game?
>>8576643Big Lust checkpoint and no Lora. See here for the setup: >>8575858. Score tags and the sampler seem to be increasing the quality.From memory the prompt was something like:'a professional portrait photo of a nun giving the vicar a blowjob at the church altar surrounded by stained glass windows' (picrel)
willie brown was here
>>8577139Thats totally random.
New drip for my OnlyNuns subscribers.
>>8576907Random redhead or super young Bridget?
>>8577176"God, i have sinned"
How come SD1.5 still makes the best skin texture by far? BigLust maybe getting close, but not consistentlyAll these newer models, XL, Flux, SD3.5 always feel like a huge step backwards to me
>>8577224SDI.5 was worked *hard*.
Anon, will you be a dear and oil your mom?
>>8577286
>>8574676Any chance you'll be able to do Michelle Trachtenberg? Or Kiernan Shipka? If you're busy it's sll good, and thanks
>>8577332yeah, should be able to this weekend.
>>8577051>>8577154this 3.5 without anything special?
>>8577051is 3.5 works in forge?
>>8573896https://files.catbox.moe/1lup7d.png
>>8577397
>>8577397Thx
>>8577224i'm impressed by this gen. I agree, the skin texture looks incredible. I feel like SD 1.5 still hasn't reached it's maximum potential.
>>8578135not sure why some of you are so stuck on old grainy shitty quality is "real". use any modern phone, take a picture of someone. their skin doesn't look like sandpaper.
>>8578158This is not an example of the sandpaper effect though, which comes from hires fix mostly btwThe skin looks actually detailed with folds, imperfections, shadows etc.Like a closeup taken by a professional photographerWhat do you think where the AI got the idea from? Photos that look like this were part of the training set
>>8578158Are you retarded? Have you ever actually seen another person up close?
>>8578193no, why?
>>8578198
>>8578205Great work! Prompt+Lora?
>>8578218biglust 1.5 with some random emilia clarke lora on civit - i'd have to check
>>8577224i have been wondering this myself. i think it's because the training set had cast a wider net and it got more pictures of people>>8578135negative prompt 1.5 is untapped meta>>8578158it's because we grew up jacking off to 480p at best
>>8577224model? promp(s)? setting? upscaler?
>>8578514cyberrealistic_v50DPM++ 2M SDE, Karras, 40 stepspositive: celeb_name,(sleeping:1.3),closeup,nude,nipples,bed room,dark room,dim lights,shadows,smirk,ponytail hairnegative: daylight,well litadetailer with face and female_breast denoising 0.55 and 0.45hires: ESRGAN_4x, 14 steps, 0.24 denoisingWithout additional negatives, makes more candid/amateur looking pic relatedAdd typical negatives deformed,disfigured etc. for more professional looking photoIf skin turns smooth try "dirty skin" in positives
>>8578546And CFG 3 (playing with lower values like that lately)If using a celeb LoRA and gens look very much different from this, lower the weight until she goes to sleep
Baking.
New thread:>>8578561>>8578561>>8578561
