Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101733170>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Kolorshttps://gokaygokay-kolors.hf.spaceNodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper>AuraFlowhttps://fal.ai/models/fal-ai/aura-flowhttps://huggingface.co/fal/AuraFlows>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
Blessed thread of frenship
>>101738336Imagine genning slop instead of memes.
aerial bombardment
Once flux training is more mature, we should somehow pool a bunch of money together and get the bigASP guy to do a finetune. Seriously that model has such a broad knowledge base for realistic human-centric, NSFW content. It's great for using as a base model to train loras on top of too.
official pixart bigma and hunyuan finetune waiting room
>>101738664no coomers should die
>>101738665how will they ever be relevant now that local dall3 with untapped potential exists?
>>101738738would be funnier if SD was missing a limb or two, but I don't suppose that's in the data
>>101738744western shemale cannot hope to beat the mandate of heaven
>>101738744Sigma team focuses on efficiency, that means they'll make a model that punches above its weight.
>>101738744It's significantly easier to finetune Bigma/Hunyuan sized models and they can 100% reach Dalle levels.
>>1017388422 more weeks
>>101738842mhm yeah totally bro I'll just trust you on that oneSD3 will be amazing too! we just can't handle immense power of the full model!
>>101738842And the biggest Flux issue is very limited style control. Once they've figured out prompts the Chink models can still shine.
>>101738873>bro people are going totally rent 8xH100s and full fine tune Flux for $100k
>>101738887pixart suck out your bone marrow and ascend to 5 star immortal realm. no cope will save you. give up.
>>101738908and yet it's been months and pixart and hunyuan have had zero movement
>>101738931my bowel movement have more soul than flux gen
>>101738931They're literally working on the next Pixart model. No movement on Sigma because 600m parameters isn't enough to get people to switch from Pony.
>>101738937>>101738943maxing out copium>2 more weeks, anon, just wait
>>101738902very nice
>>101738951mass western suicide on september
>>101738937hecking based I also love SD3 & Joe Biden
>>101738951I'll be using my 1.3B model, I promise. :)
Bigma status?
chang on suicide watch
pinnacle of flux genner creativity: 1girl realistic please also donald trump
>>101738379woah my Slime Girls made it into the OPfeels good man
>>101738986woaahhhh 1girl anime big benis!111
>>101738664He trained 5 days sdxl on 8xh100 ~3500$. This does not include all the tests. So about ~5000$ for bigASP.For flux you can estimate 20-30 times as much and nobody knows if it will work.But nobody will pay a cent because now that we have trainingsscript all the finetunes come for free.Tranny fag users have promised me that
>>101739010hatsune miku holding a sign reading "nigger nigger nigger nigger"
>>101738974still in the land of the cursedalso the latents are going to take forever for 512pxeach shard of 20k images is an hourand there's hundreds of shards
>>101739027all your shit can't even do that? lmoaing@you're life.
>>101739040i outlive you xister
>>101738744flux looks like shit. if bigma can actually generate cool styles in 8b i'd use it more than flux
>pixart best-art, now go, tell the wolrd
>>101739027Uh uhm y'alls gender reveal when?
>>101739057september when i reveal myself to be a pixartsexual
>>101739056yes, Pixart does better fine art than Flux
>>101739052>>101739053>I-I-I don't even want prompt comprehension! I swear!>y-yeah, base bigyuan diffusion looks like shit, but we just need to finetune it!!
how can I change the CFG value in the flux workflow?none of the nodes have the setting
>>101739070>fine artlmfao
>>101739087>nogen as alwaysgive your best Flux fine art
>>101739032sounds unoptimized. are you writing individual files? also use torch.compile or some other acceleration
>>101739053It's overcooked or just RlHF'd to look too much like dreamshaper and it's not good at manga gens. I think this is one big reason it sucks aesthetically. The major setback of RLHF is if authors are not careful then this happens. There's that Flux look, and it's basically the same look that all SD 1.5 and XL has.
>>101739077>I-I-I don't even want prompt comprehension! I swear!the xister reveals xerself to not know what xhe's talking about. aiiiiieeeeeeee not good not good.
>>101739084Double click the background and type cfg in the search box maybe. Or steal a workflow that has it. That's what I did.
Are any of the new non-SD models usable on non comfy guis?
>>101739128nop
>>101739084"FluxGuidance" nodedev supports guidance factor but schnell does notexample workflows at:https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-dev
>>101738998congrats anon, the slime girls are very cute>>101739044woah
>>101739093:3
>>101739133Thanks. Sad.
How good is Flux at making cute anime catgirls? Did it surpass DALL-E?
>>101739109The fact that they used synthetic slop to refine their model is concerning, and that helps explain why it's so bad in many cases.
>>101739070they will never be able to counter this
>>101739151it's pretty good at anime gens from what i've seen
>>101739102>nogen while crying about nogenthis is your brain on copium
>>101739113>meanwhile he has literally nothing to back up his two more weeks claimsVERY SAD
>>101739174what two more weeks claim
>>101739172>he calls this art
>>101739172this my brain on lethal dose of ai sloppa. i die, AAAAAAAAAA!!!!!!
>>101739142
I've replaced my 15 year old pc: a new ryzen 7900,64 gb of ram and 4060ti 16gb I've just entered the dynamic rabbithole of stable diffusion and pony,and now we are gifted this new powerful flux modelwe truly live in interesting times... it's like magic. I can't wait to try and experiment new models
>>101739191>IT WILL BE GOOD I SWEAR!!!!!1111
>>101739199pretty sure I didn't call that anything, chang
>>101738942I like this style!
>>101739172noooo the heckin style is LE BAD0/10 I can't masturbate my tiny peanus weanus to this!>>101739216why would you get a 7900
>>101739104Yeah it's writing cache files, on the 3090 it's 6 images/s to do the VAE and T5 caches.
>>101739222>incoherent babblingmind broken already? aiiieeee western man too weak! too weak!!
>>101738379Holy mother of slop
>>101739241>0/10 I can't masturbate my tiny peanus weanus to this!I can
>>101739241i masturbating to you
>>101739151You can get good results with anime, i'll try catgirls too,lolAs far as i've experimented, she knows Reimu Hakurei,Remilia Scarlet, but it's not always coherent to prompt,sometimes draws another character
where can I download flux at? Huggingface wants me to sign in and agree to shit but I'm not doing that
>>101739277You call that anime? that's not anime! It's just BAD! and SAD!
>>101739277I can be a struggle but we'll get there. I believe in the potential
Uhm uhhh y'alls love Biden & SD3 hecking chuds
>>10173924112 cores at 65 watts, i want an efficient pcAlso consider that i was still gaming fine with my old 1st gen i7 870 and rx 470 8gb
>>101739316you loves estrogen pills i handmade at china factory. i added nice surprise!
>>101739317yeah, but do you even need it like ever?I only got one because of a pricing error
has anyone compared sampler/scheduler combos in flux dev and schnell? some of them are slow as shit
>>101739341do not engage just report and move on
>>101739249fuck yes anon I want this prompt
>>101739243individual files will slow it down a lot, you should modify it to write the batch or look into webdataset. what are you using, something you made yourself or something built into the trainer?
>>101739359this is really cute. Moar!
>>101739151Everyone here is blind. No, it's not that good at anime, at least not stylistically. Dalle will always have more kino because it's not made to look like dreamshaper.
>>101739370I essentially took what they use for caching for Sigma and built it into the dataloader, it just simplifies it by putting the npy and npz in a cache folder.
>>101739358Uhm y'all?
>>101739384
>>101739134ok that worked for me so whats a good CFG value?
DEAR MOTHER OF GOD
>>101739409kek, give him a wand :D
>>101739410From what i've heard cfg and flux guidance are different things, because on flux cfg is stuck on 1 unless some trick related to dynamic guidance
>>1017394103.5 is default. start there and play around
>>101739421
>>101739458That's awesome! Thank you
>>101739466brap
>>101739458lol amazing
>>101739484nice mosiac ass
>>101739385Yeah, the anime artstyle variety you can get with Dalle is still unmatched, sadly. Flux looks too "realistic" to me.Though, I've yet to see some "anime watercolor" or "manga style" examples from Flux.
>>101739542A shame, really. You can do all this with Hunyuan and it works out nicely but it's got only a fraction of Flux's prompt adherence.
Wasn't sdxl trash at anime (even worse than flux) and also censored against nsfw when it released? This gives me hope for something like a flux-based equivalent of pony down the line
young me learning to appreciate fine art
>>101739637Throwback to SDXL 1.0 release
the lora training has begunhttps://x.com/ostrisai/status/1820462674230059328
>>101739542while I have no idea bout dalles anime style, I agree that I am not happy with the standard anime style that flux delivers, I have yet to find a way to prompt it to get to a style I like, with SDXL that was easy to get a style lora you like and it will be consistent . I have to feeling the anime dataset in FLUX is not curated well, sure it has dragon ball and hatsune miku, but beyond that its abit meager
>>101737857The lattice you see on here is actually what flux does when it's cooked. It's not you getting closer to your style, unfortunately. Your guidance will be too low. >>101739673Lora training began yesterday. Took him a day to complete, which is interesting
>>101739637>>101739672 (Me)anime one
>>101739672My first SDXL gen
Thanks to this trick we can use CFG, negative prompt, and niggerify Miku kekhttps://new.reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
I heard this place is better than sdgdoes d**o post here?
don't give me that look
Some good new swifties coming out of flux
>>101739739no he doesn't, thank god
>>101739739>does d**o post here?newp, but chang does
>>101739774I like Swifty as a chubby sexy midwest hotwife.
>>101739739please do not mention namefags
chibis are kinda nice in flux tho
>>101739729This cooks your images in most cases by the way. It's a work around, not a solution.
what's up with the split?
>>101739792That doesn't look like Taylor Swift.
>>101739815sd lost
>>101739808I'm pretty sure the parameters on DynamicThresholdingFull can be improved and make it look close enough to cfg = 1, need some more testing I think
>>101739815/sdg/ is a chat room for avatar fags>Hellooooo cumbot>Hiiii fago, how was your day>Oh you know me, hey has anyone seen dikdik today?
>>101739815think of ldg as lmg to sdg as aicg
>>101739843holy kek
>>101739833And it will still be a work around. We will need to bake in negatives with training, if it's even possible to at all.
Ah, there's my Piggy Swift.
>>101739681>SDXLYeah, no thanks, that looks a bit too much like slop to me.I invite anons stuck with XL to take a look at https://desuarchive.org/_/search/boards/g.desu.meta/filename/HunyuanDiT/page/1/It is superior over SDXL with multiple subjects, hands, etc.. and is a lot closer to Niji/MJ. Those are gens I made across random tests and the prompts are in the filename. This is from the base model alone.They do not contain all my gens, some of them are in responses like https://desuarchive.org/g/thread/101092621/#q101104908Training a LoRa with it simple.
>>101739024>now that we have trainingsscript all the finetunes come for freewhat
>>101739867I found a workflow that works with negative prompt and let the CFG at 1 (it also use DynamicThresholding), maybe that's what you wantedhttps://civitai.com/models/625042/efficient-flux-w-negative-prompt
>>101739681Yeah but remember, you didn't have LoRAs SDXL came out, that took a few weeks. Or maybe you did, I don't know how new you are.
>>101739923I don't want anything, I am saying threshold scaling is a bandaid to a baked in problem. It's like trying to add more flour to a cake after you already cooked it. You can kind of do it but the results aren't good.
>>101739896in the rare case I crave anime gens, I will look into it>>101739928ya memberberries
>>101738986trump nigger jews obama swatiskasdid I forget something?
>>101739996>upset that 4chan used AI to make some swastika trump jew nigger images, which isn't even the majority of images posted herego back
all flux gens look the same
>>101740039WHERE DO I PURCHASE?!
>>101739923This is 3x slower and produces results that look worse.
>>101740077can you show me those pictures "before" and "after"? I wanna see the disaster kek
>>101739681Maybe it's because they scrapped many characters and artists.
>>101739681it has not beeing trained on enough styles, characters and celebrities yeah, the first big finetune would be to fix that and add new trivia to the model imo, because it's a great base model on everything else
before I give flux a whirl, what was that thing I read somewhat recently about ComfyUI sending telemetry to third parties
>>101740077>results that look worse.like even worse than the "old" workflow on reddit?
>>101740095I really wish they would answer questions about the data set, but Black Forest anon yesterday just a political "idk", guess we have to it the hard way and do broad tests for artists.
wtf comfy crashed and now it won't load
>>101739542No way you are serious. Dall-e seems to have a reduced number of styles, But the worse thing is the defining texture that it gives to everything I can't understand how people can stand it.With SDXL you can access hundreds if not thousands or artistic styles (better and worse) and mix them to your heart's content. if you get tired of exploring it's your own fault. Flux is limited in style but it's really good at prompt understanding, it's creative and it gives very nice compositions, then you can img2img gens to add style with a SDXL based model. It's not perfect and sometimes SDXL changes too much, but there are cases where the changes are lucky and the picture turns out even better.
>>101740130it's open sourcethe code is right there
my fellow americans
>>101740151it wasn't phoning home
>>101739681it even can't do picasso, which is insane that's not even a "copyrighted style", but it can do Trump and Swift fine even though it's way more controvertial to have than than fucking Picasso, I don't really get their motive behind that
>>101740151got pickled
>>101740204I can't deny the texture one can get are really fucking cool
>>101740165did you ask for a bimbo or just described her?
>>101740216both
>>101740168>>101740183Imagine I am completely retarded. What does this mean? Or more importantly, how do I fix it?
When you type "nude", and the model doesn't give a shit, does it mean the training data for "nude" images doesn't exist, or they just stripped the word "nude" from the model itself?
>>101740207Reminds me a bit of the game Monkey Island
>>101740230Both. Or either.
>>101740230they misled the model on believing that "nude" needs "topless people with panties", that's even worse than having the model not understanding nude because it means that the model needs to "unlearn" that shit
>>101740267Alright!
>>101740229>>101740151Never mind I fixed it by updating pytorch
>>101740177Picasso works fine for me:>cubism style, woman, by Picassowell maybe abit to perfect for late period Picasso, but early period Picasso definitely painted like t his
>>101740230>>101740272It's neither of these things
>>101740284we're just being stupid anyway
>>101740177Try a different artist. There are reasons for Picasso being problematic to how AI works:Picasso didn't have a single a style, he was a genius that excelled in drawing since he was a very young. Also Cubism is just too alien to how SD works, because it's figurative but it's created by the artist in a very different way.. Maybe a model guided by an LLM can give you the Picasso style you are thinking off, but as a token I understand how it works poorly.>>101740294That's cartoonish but it's not a cubist picture at all
>>101740294go for "hatsune miku in Pablo Picasso style" and it doesn't work anymore
>>101740301I encountered more women with panties than I have not though, the model is confused on what it should take
>>101740301Then how did they do it?
>>101740307>That's cartoonish but it's not a cubist picture at allearly cubism was very cartoonish pic related>>101740320cause hatsune miku comes in with baked in anime style
Sorry for my autism but does anyone have a high res version of this? Im trying to find the original source. Or are most ai pics low res?
>>101740338>cause hatsune miku comes in with baked in anime styleand the model still should be able to transform hatsune miku into a cubistic style, like every other models we had before, flux can't do that
>>101740350booba
I feel a strong urge to scam buyfags.
>>101740350you can hi res it yourself
>>101740332>>101740335My best guess, and keep in mind this is only a guess, is they are adding addendums to your prompts to make it "safe".
>>101740360With what? Waifu2x?
What's the actual difference between the pro and dev model outside of the pro being behind closed doors?Is the pro as brainwashed as the dev one?
>>101739923I tried it and the neg prompt works yeah, how did he manag to make it work at "cfg = 1"?
>>101740363TWU
>>101740361not possible to do that if you're just running model weights locally
>>101740375I would suspect even moreso, but I dunno because I've never paid for an online service in my life
>>101740375dev and schnell are distilled versions of pro, so finetuning is abit trickier
>>101740375haven't tried pro but it's supposed to benchmark significantly better than dev
>>101740389Love this movie look. Please do more!
>>101740301brap
>>101740375My guess is it's like DALLE, fully uncensored but with a moderation type rewriting your prompts and stopping anything deemed non appropriate.
>>101740152Nta but you can't do anything like https://desuarchive.org/g/thread/100695227/#100702221 with either Flux or XL. With Dalle and Hunyuan you can. (XL in particular does not follow styles as well as a DiT model). You'd need a finetune for Flux, but a base model is always going to have more variety and flexibility over a finetune (see how all XL finetunes look the same), and Flux already has a very similar style to those XL finetunes... Also that example is just one of many things you can't do with XL. Dalle may be limited in style control as well, but it has an aesthetic set of default anime styles.As for img2img that is a possibility but you're taking an extra step and the goal should be to reduce them. If it doesn't know the artist the subjects may be off depending on how different the styles are.
>>101740375I tried both and I didn't notice much difference, they really gave us a great model that is really close to their API ones, so fucking based I kneel
>>101739808Mess with the settings. That guy suggested interpolate_phi at 0.8 but that cooked my gens too. Try 0.5 instead.
>>101740361>>101740388Yeah, which makes me still wonder, what the hell did they do to get the model to "forget" what some concepts are.
Which version of Flux is better for the 4080 Super, Dev or Schnell?
>>101740442SchnellAnything below a 3090/4090 VRAM wise is schnell.
strippergram
>>101740350
>>101740442fp 8 dev
>>101740441it's not forgotten, vllms will caption half-naked as "nude", they just were extremely good in ensuring there are no genitals in the dataset, as far as I can tell there's little to no genitals, one time I saw really, really blonde pubic hair but it could've been panties
>>101740449absolute lies. Takes me ~2mins to gen and upscale w/ sampling on a 4080
I call skill issues. >cubism style, Hatsune Miku, by Pablo Picasso, abstract geometric face>neg: anime realisticguidance 3, cfg 2, mimic 1
>>101740418I wonder if they'll follow SAI and make sure all future models are worse.
>>101740383led to some interesting outcomes
>>101740381>how did he manag to make it work at "cfg = 1"?Looks like PerpNegGuider can make you able to use negative prompt without any cfghttps://ku-cvlab.github.io/Perturbed-Attention-Guidance/>Perturbed-Attention Guidance significantly enhances the sample quality of diffusion models without requiring external conditions, such as class labels or text prompts, or additional training. This proves particularly valuable in unconditional generation settings, where classifier-free guidance (CFG) is inapplicable
>>101740478that's not cubistic at all, her eyes, mouth, ears aren't all over the place, do you know what cubism means?
>>101740449I go from a minute twenty to about a minute on schnell which doesn't seem worth it for the quality loss. 3060 here 1024x1024
MAKE IT DO INFERENCE FASTER YOU FUCKING RETARDS PLEASE
>>101740512>>101740528lol
>>101739923>Perp Nigger Guider
>>101739808MmmmmCrispy
>>101740462Thanks anon.So basically a mix of curated dataset, and using the wrong words for descriptions.I guess they also did that with artists/characters.Sounds like a lost cause.
>>101740137Reddit examples are above. >>101739808No negatives followed my prompt more closely. Anons workflow removed he horns for some reason. Quality is the same but adherence is worse, and it's 3x slower.Reddit workflow is cookedSD3 for keks.
>>101740529thats wrong cubism is a wide array of non realistic depictions that go from simple geometrization to total absurd collages of forms, Picasso only painted from 1909-1919 in that style and he experimented from simple gemetric phases on faces tot total abstraction .. what you want is extreme cubism that you all have in your mind when you think of picasso, but I am sure FLUX is trained on the whole of Picassos work .. but Okay ill try to make a cube block Miku if you want it
>>101739466>>101739659quite edgily funny
>>101740585>but Okay ill try to make a cube block Miku if you want itgo for it anon, I'm waiting, and you need to make it work with only "Hatsune miku in the style of Picasso" because picasso only made cubism in the "fucked up eyes, noses, ears..." sense
>>101740566yes they most likely used a name filter on the dataset and vllms don't really know names either, Florence labels most celebrity pictures as "man" or "woman", the exception are people like Trump which it seems to do right. Same with pop culture, either it just says cartoon or animated and worse sometimes incorrectly labels the character. I have a feeling they have zero raw captions in the dataset either, probably all AI generated.
>>101740576>sd only one in it/s Sad.
>>101740576>Quality is the same but adherence is worse, and it's 3x slower.by "adherence" you mean even only positive prompt has worse adherence than the "normal" setting?
>>101740603>because picasso only made cubism in the "fucked up eyes, noses, ears..." sensethats so wrong I can't
>>101740616nta but how do you even make up for thatI don't see how you can train anything back with so much missing.
>>101740630nta but having the script invert the unit is absolutely retarded
>>101740640give us your workflow anon, looks like you managed to make Miku into the cubism realm
>>101740640not bad kinda got the asymmetric eyes
>>101740644SDXL was dogshit for porn but led to naiv3 and pony, so it can be done
>>101740637Yes. Which is pointless. The positive prompt is stuff I want to see.But the main issue is it is 3x slower for very little benefit.There's definitely use cases, but to just gen and shitpost it's only worse. You aren't going to be able to use negatives to generate porn with Flux.
>>101740616>I have a feeling they have zero raw captions in the dataset either, probably all AI generated.That's probably not true, or else it would mean that if you put a celebrity name, it would give you someone completely different, the model kinda knows how the celebrities look like, it's close but not quite, imo I think the model was just undertrained, Trump looks fine because there's probably gazillions of pictures of him, so the dataset has enough Trump pictures to make it work
>>101740684>You aren't going to be able to use negatives to generate porn with Flux.don't doom like that anon, a few days ago people said it was impossible to finetune "distilled" models and they got proved wrong, more researsh will be done on making neg prompt work without CFG
>>101740644Probably takes a few epochs with a large dataset to fix it, it depends on how flexible the model is to learning new concepts, SDXL with insanely difficult to fix which is why porn models took months to get good and it really wasn't until Pony that it properly did anatomically correct vaginas.
>>101740667Was it as curated as flux though?I wasn't there when it was out, last time I played with this was 1.5 pickle era.
Only buying a new GPU when Flux can do pornBy that time GPUs with a sufficient amount of VRAM will be cheaper
>>101740640how did you made that work? I thought flux had no idea what Picasso style looked like
>>101740705>finetune "distilled" modelsWhat makes a "distilled" model technically non finetunable ?What even is a distilled model?
>>101740744>By that time GPUs with a sufficient amount of VRAM will be cheapershould we tell him?
>>101740686You can test it yourself with an only vllm captionercase and point with my own captions
>>101740705I am sure you can fine tune porn in, but you definitely can't negative the panties off your women
>>101740752if I understand this shit well... "distilled" means that they compressed the model so that it's faster in the inference, because a 12b model is a huge boi, if we had the "normal" one it would take ages to make a single picture on a 3090
>>101740769if you put enough nudes pictures on the finetune, the model will unlearn the fact "nude" women have panties
>>101740750Picasso style is not what you think it is, he had many many styles, you need to help flux abit to get into Picassos cubism.ya here: >https://files.catbox.moe/cos8j3.png
>>101740722Imo SDXL was better at artstyles. Could be a skill issue on my part but flux is resistant to some of my favorites. It's great at some stuff like 3d Disney art though.The best thing is the prompt adherence. When I ask for underwater McDonald's that's what I get.
i've been shopping around for a vid card upgrade so I can start making sdxl loras, anyone know a good cheap one? you need what 12gb min right
>>101740791the face is still too clean though, Picasso made more cubism paintings with displaced eyes and shit than not, the model should know that
>>101740755>implying 24GB VRAM GPUs will never be cheaper than 2000 leafbuxA decade ago 8GB was top end
>>101740752>What makes a "distilled" model technically non finetunable ?Nothing, someone already trained a lora. >What even is a distilled model?A distilled model in image generation is a smaller, more efficient version of a larger model, trained through knowledge distillation to retain similar performance while using fewer resources.
>>101740803Krupp coastal cannon
>>101740790Yes, that's what I said. What I am saying is with the current base model flux, you can't use negatives to make nudes. I know that's why people want negatives, but it won't work.
>>101740796>The best thing is the prompt adherence. When I ask for underwater McDonald's that's what I get.Yeah Flux is the first time I feel like I can gen images and have it actually understanding my fucking prompt. All the previous models I've used felt like it would just take a few simple concepts in my prompt and completely ignore the rest. Being able to describe stuff like "the image is split into two halves, X on the left, Y on right" and it actually works is so refreshing.
>>101740644Stability did with SD3, they trained first half (billions of images or so) on AI tags, then second half (I assume containing most fine art, celebrity, etc... references) on web scraped data (just leaving it as is). The best solution is to use both and prompt the LLM to say "art by" and/or link the data.
>>101740771>>101740819I see, thanks.
>>101740760I think your prompt is easy for the model to do because there's only one movie poster of Skyfall, so you're just asking it to redraw the same picture it has learned before, that can't be comparable to a celebrity that has millions of different ways of photos of him, a model truly learned something when it's able to make something completely new out of that character
>>101740814hard to replicate naive style when it's not in the training set
>>101740847You missed my point, did it say Daniel Craig or did it say "man".
>>101740843so theoretically either kind of text encoder(?) would work?
>>101740831oh, but negatives isn't just to remove panties, it has a lot of possibilities, you can remove colors, logos, realism, drawings, basically it helps the model more by saying to it to not use its biases and go for something newer
ever since I fucked around with the CFG shit the generating became a lot slower.even after I changed it back to the standard workflow.what this means? restarting also didnt fix italso what it means when it says:>UserWarning: 1Torch was not compiled with flash attention.... ?pls halp
>>101740816>implying the dollar holds the same value as a decade agoand how much was a big mac?
>>101740760can your vllm captioner recognise porn?Any xxx fine tuning will need to use a vllm to caption the data set, so I am curious
>>101740853you also missed my point, the model knows how the man looks like because if you ask for a poster movie of skyfall, it has to be this man and no one else, like I said, you just asked the model to draw something he already saw in its training, that's not hard to do
>>101740814yes, but its still hatsune miku, she is strongly symmetric I can add more tags to make he rmore abstract maybe lower the guidance further, the basic idea is achieved tho, the super abstract Picasso works are surrealism, not cubism
>>101740872Yes anon, 1 GB of ram takes the same percentage off your paycheck as 15 years ago
>>101740709It's not just nudity/porn though, it's known characters, styles, artists, people, etc.That looks like an insane work to even get started.
>>101740882>yes, but its still hatsune miku, she is strongly symmetricevery human is "strongly symetric" anon... that's why cubism is what it is, it breaks that symetric thing out of human faces
>>101740875A typical porn caption looks like:>a young woman lying on her stomach on a wooden floor. She is naked, with her legs spread apart and her hands covering her vagina. The woman has blonde hair and is looking directly at the camera with a seductive expression on her face. The background is blurred, but it appears to be a room with a wooden cabinet and a white wall.But I trust the Booru tags to cover more pornographic concepts. The T5 is very good at mixing regular captions with specific poses.
>>101740892Doomers will doom anyway anon, why care.I'll be really happy when all these 3090/4090 will be in the used market with the new gpu in the horizon.
>>101740882is it possible to get it without that peeling brushtrokes look? I noticed it REALLY likes to add this sort of grunge worn-out punk feel to the painting as if they've been rotting away or something, even on this cranked cfg one
>>101740760Interesting, so the images are there, just mislabeled on purpose.
>>101739334Cool, what prooompt?
>>101740791your settings on the dynamic threshold are different to the one on reddit, you found something better?
>>101740914you are talking about late stage analytical cubism, picasso was part of early cubism..ah read it up your self>https://www.cubismartwork.com/analytical-cubism/face-decomposition/>>101740941cause the paintings look like that .. go to a musuem and look at em,
>>101740951>prooomptI don't respond to faggots. Try again.
>>101740949it's not "on purpose" I don't think vllms are very good at captioning pop culture, but I do think they removed names in post processing on purpose, even Florence on occasion does the correct name but I think they're accurate enough to trust to give the right now every time anyways so it's probably better to err on "woman" than give the wrong woman's name.
>>101739535Have you been able to get severed or deformed limbs on people to any degree? With ideogram I never succeeded. I am also rn unsuccessful with Flux.
>>101740964yes constant is better than sawtooth, thats just nonsense
>>101740941Try more modifiers like "pristine condition" maybe?
>>101740986Thanks, anon. I always knew I was straight.
>>101740996I'm not a big fan of the "AD" parameter though, it oversaturate the picture, try "STD" instead
>>101740996>extreme cubism style, Hatsune Miku, by Pablo Picasso, abstract geometric face, eyes and mouth at the wrong placedesu the model should be able to make your picture with just "Pablo Picasso style, cubism", if you can't find a single seed that has displaced eyes then something's wrong with flux
>>101740978no they fucking don't, it's applying a rusting metal effect to paint. just look at this shit >>101740996 that's not how paint deteriorates. just look at the background of the mario one, it's rusting. for some reason it associates paint with metal/rust
Nobody expects the curved barrel
>>101740902Flux definitely knows these things though. It's for sure in the data set. You can get these things through descriptive prompting without naming the character.
>>101741012FLUX looses saturation when you raise CFG, thereore I found AD better .. but your point is valid elsewhere>>101741030thats just your opinion man, I prefer a model thats exact, not just makes what the majority thinks is correct
>>101741060>thats just your opinion man, I prefer a model thats exact,but the "exact" thing is that Picasso made pictures with displaced eyes on his cubism paintings, that's just facts >>101740814
>>101741060Ideal would be a model that generates what the majority finds correct by default, but also correctly obeys any further detail you specify
>>101741038It already gives Sonic the Hedgehog if you prompt with "Sonic the Hedgehog". Can you give an example of it giving the image of a thing when you give it a description of the thing where it cannot also give the image when prompted with the name of the thing?
Picasso lora when?
>>101741078that's why a stylise slider is needed
>>101741038Yeah.I'm copy pasting wiki description and it works.Insane they got rid of so many known pop culture characters.What's even the point of these models.
So what can I do when I get the message that there are no Cuda GPU's available? I'm new to this AI stuff its fun as hell as I'm making images for a screen play I want to realize, but now I'm stuck. I'm using Fooocus if that helps at all. Thanks. I really want this fixed so I can continue.
>>101740870Because CFG bigger than 1 takes double process
>>101741038>>101741083I don't know what it doesn't know
>>101741106>I don't know what it doesn't knowit doesn't know a lot of stuff, can't make an accurate picture of Messi, Ronaldo, Djokovic, doesn't know a single character from Love Live, the list is long...
>>101741074the ones you know .. pic related, picassos first cubist self portrait, one of his most iconic cubism pictures
>>101741135Most gacha characters are also unknown to it, despite some having a giant fan art following. I wonder if autistically describing them would bring them back.
>>101741131Looks like a mix between factory town and factorio. What kind of prompt gets you the neat conveyor belts that actually connect?The yellow triangles look like out of power symbols from facotrio too.
>>101741106Pick some celebrity that should be in a huge number of pictures online but that it doesn't know how to generate when prompted with the celebrity's name.
>>101741154My point is that picasso DID MADE cubism paintings with displaced eyes, and Flux is pretending he never did anything like that, you understand what I'm saying now bucko?
>>101741173The physical traits of celebrities are much harder to describe than cartoons. The celebrities are definitely in the data set, as sometimes when prompting for Wonder Woman I get Gal Gadot, but if I prompt Gal Gadot, I get a random woman.
>>101741173Sometime it knows them, at least the dead ones. Marilyn Monroe wasn't scrubbed for example.
>>101741201Lynda Carter is better wtf
>>101741103https://github.com/lllyasviel/Fooocus/issues/1189https://github.com/lllyasviel/Fooocus/issues/1681
image quality significantly raised but also slop levels at an all time high
>>101741201Managed to gacha an example>>101741220I wonder if I specify 70s I will get her.
>>101741168https://files.catbox.moe/1pii6t.pngProbably>a colossal factory complex surrounded by smaller outlying buildings interconnected by numerous moving conveyer belts laden with raw materials shifting ceaselessly towards their destination within the main structure.
>>101741175>with displaced eyesyou are talking about surrealism, not cubism, and yes surrealism can use cubist elements
>>101741244write "Picasso cubism" on google image and you'll get pictures with displaced eyes, ears and so on
Does anyone know what kind of prompt can enforce the analog/VHS look? Even after prompting a lot of vintage analog related stuff I still cannot get it right, it always looks too clean
>>101741244I prefer Picassos pure surrealism, pretty whack stuff
>>101741225 (me)>>101741261 is soul actually
>>101741233>>101741201Example 2, not Gal Gadot
>>101741252forever delitilized
>>101741258picrel
>>101741270forgot pic
>>101738751WHY CAN'T IT DO DEFORMITY?!!!
>>101741252sex without consent
>>101741269turns out artist styles add soul as opposed to whatever the average of all the dataset produces. not sure how many made it though, certainly not as many as i would've liked. also seems that its a war between concept and style, hard to get complicated interacting prompts in a style
>>101741285you have SD3 for that
>>101741038There is no doubt stuff is in the dataset. But since it's untagged it might as well be a guessing game, and unfortunately the same thing happened to artists. Their names got lost. Generic descriptions can only do so much in describing what something is.
>>101741299So they made their own model worse by design.
>>101741290explain, I'm new. Flux won't give me deformed limbs (yet?)
>>101741312I dunno, not a lot of amputees in the training data?it'll probably work for drawings or characters like MK's shiva or the man spider
>>101741312Pirate hook or peg leg might work. Maybe prosthetic
>>101741352Put this in the collage for christs sake
>>101741221Crap I knew you would post this. I cant make heads or tails of it.
>>101741352kek
>>101741310The only thing we can hope is that they fix it with next iteration of the model, if you scroll down here-https://blackforestlabs.ai/to pic of the red moon then you see that it was actually prompted with artists, so it's not clear whether or not this was intentional.
>>101741400And the 1 in FLUX.1 probably means version 1? But this is all speculation at this point.
>>101741396lol Munch! need to try him in Flux>>101741356>surrealism style, Hatsune Miku, by Pablo Picassoguidance 2
>>101741400I think they went too far by design, and we won't get these concepts back in a future iteration.But who knows.
baking hol' up
RIP ldg
>>101741458thanks robert!
Bread, seeing as the other guy didn't >>101741569>>101741569>>101741569
>>101738568man this is freakin awesome how do I learn this power??
