Previous /sdg/ thread : >>101697538 >Beginner UI local installEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio >Local installAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUISD.Next: https://github.com/vladmandic/automaticAMD GPU: https://rentry.org/sdg-link#amd-gpuIntel GPU: https://rentry.org/sdg-link#intel-gpu >Use a VAE if your images look washed outhttps://rentry.org/sdvae >Run cloud hosted instancehttps://rentry.org/sdg-link#run-cloud-hosted-instance >Try online without registrationsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-mediumtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest >Models, LoRAs & upscalinghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://openmodeldb.info >Black Forest Labs: Fluxhttps://huggingface.co/black-forest-labs/FLUX.1-schnellhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux >Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd >View and submit GPU performance datahttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance >Share image prompt info4chan removes prompt info from images, share them with the following guide/site...https://rentry.org/hdgcbhttps://catbox.moe >Discord6wUwtcJsr2 >Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>>101700396Curt Cobain version please
rip porn models
Where are the controlnets? IP adaptor, xinsir unified controlnet, finetunes? DEADONARRIVAL
scribble is hard
It's no censorship DALL-E, open source won.
>>101700504you can take that and do openpose on it, and it should give the pose info and then you prompt whatever you want. scribble is tricky but you can get a lot of cool outputs with it, and then do other stuff with that.
>>101700485it would take an ungodly amount of compute to finetune, but it's not impossiblepony freaks have already died on this hill, they have nothing left to lose
>>101700519and their work yielded the best models for making anime characters, even if it's good at one thing it can also be good at another
>>101700519>it would take an ungodly amount of computeare you niggers seriously that poor and cant rent 4xA100s for a day or two to get decent results
just tried out flux for the first time and it's crazy good for a base modelthis is a very basic gen but the clearly readable hiragana is a million times better than you get with any SDXL model...but idk what to do with it until there's shit like ipadapter and controlnet
>>101700544I don't have hundreds of terabytes of training datasets
>>101700545you exploit the good text and prompt understanding to make memes
>>101700545anything you want, that's the beauty of itstop thinking in 1girl, hands on hips terms and go wild
>>101700565better:
>>101700485They gave us the drug before taking it back from us, what a toxic poison gift lmaoWhat is the reason we can't ? I mean if you have enough GPU you should be able to finetune it right ? Or did they use some ancient black magic to protect their virtual data across the globe ?>picrel: miku live reaction
>>101700484I tried kurt cobain as a muppet
>>101700582there is absolutely nothing stopping you from inpainting gens or img2img in ponyXL which excels at porn
>>101700591That doesn't explain why we can't stillWhy especially this model ?For LLM, we can train big model across multiple GPU or load them in 8bit/4bit for finetune to use less VRAM, so why can't we here ? What is the catch ?
>>101700583Very nice
>>101700608I wouldn't worry about it, there's a lot of buzz around it so we'll find out pretty soon
>>101700608it's day 1, I guarantee you people will add stuff like controlnets or what have you. but even as-is, it's very very good at depicting a range of prompts/styles. The text depiction and range of text is also amazing (it can do cursive even)
>>101700608>some obvious tranny with a fake ass Japanese name says you can't fine tune the model in a random screenshot of a discord chat>Suddenly call into question the well established fact that all models up until now have been trainable.Think about it.
>>101700617I just want a little bit more lewd/erotica, not even porn. It's so frustrating. But I guess you're right. I mean a small LoRA would do the trick I'm pretty sure.
Can someone drop a catbox using flux? I cannot get the fucking UI to work and I wanna kill myself.
>>101700623>>101700625Yeah y'all right we can only wait for now
>>101700485literally who?
will try to sleep, though i stayed up way too late and will struggle to fall asleep.today turned into clip caping random pics and pasting random tokens into the prompt, and... all bad results!goonite
the stuff that dalle will reject is now fineOPEN SOURCE BABY
I take it a VRAMlet like myself won't ever be able to run flux, right?
Guys whats the most cheapest most bargain bin card that can run current SDXL models, no im not looking for middling or average performance I want an absolute bargain bin poorfag card that can just barley cross the threshold to generate images, Im asking for science I guess.
>>101700705if you can run other FP8 shit you can run this. Just not at it's full glory (unless you don't mind waiting 30 minutes per image)
>>101700721Any card made after 2016 with 4 GB supporting a high enough cuda version for your UI of choice. Use Forge or SD.Next or Fooocus, they should have enough optimizations to run it fine as long as you have 16 GB RAMSee https://developer.nvidia.com/cuda-gpus
>>101700681Goodnight, sleep well!!I think it's about time for me to sleep too hahaGoodnight, anons! :]
What celebrities does Flux know other than Trump and Biden?Joe Rogan isn't one of them
>>101700761Alright thanks, I'll take a look & see what has later CUDA ver. and at least 4 GBs or VRAM, all I needed to know thanks again.
>>101700782I'm personally using a Quadro P2000 with SD.Forge UI and I get acceptable times on sdxl/pony, with 16 GB RAM, of which 3/4 are occupied by the genning (depending on resolution)
lmao it even knows styles like 1940s comics, this is greatfor this prompt I didn't specify dialogue though.
>>101700846That's an actual Joe Biden quote too.
>>101700631be more creative with your prompt if you just want eroticaalso good morning anons
>>101700782you want 8 at the bare minimum, better 10 or more
>>101700867Imagen doesn't normally draw out the creative types
specifically mentioning white speech bubble did the trick.
>>101700846>>101700952neat
lmao, this one worked well>Batman punching Joe Biden in the face with his fist in the oval office, Joe Biden looks confused and not sure where he is, Batman is smiling and saying "get out, old man" in a white speech bubble, Joe Biden is saying "Hello, Spiderman." in a white speech bubble, the image is in the style of comics from the 1940s
>>101700976>>101700976we have come so far yet have a long way to go on hands
>>101700987we've come very far with ponyXL/SDXL and now this
>>101700517now I'm trying 3D mannequin image + scribble and it worksmannequin is good when you like the body
>>101700996neat, theres something called openpose editor where you can make a pose for controlnet, but that looks even more detailed
>>101700992XL is a pile of garbage compared even with Pony
>>101701021well pony is hyper trained on everything so it works great for anime or even realism with the right modelautismmix-confetti is top tier for anime gens for example.
>Joe Biden is in the oval office, Joe Biden looks confused and not sure where he is. Joe Biden is watching a TV with CNN on the television. The television has a chyron saying "why you are racist" with a black news reporter holding a mic. The image is in the style of comics from the 1940sthe prompt understanding is really good. i'll still be using sdxl/ponyXL though for characters and controlnets, but this is a very good open source model.
>>101701028Pony is the most fine tuned model out there, and it's getting mogged by a foundational model. SAI is dead. There's no way SD3.1 can compete. I feel bad for whomever is still giving them money.
>>101700770Take it easy dude!
>>101701044all sd3 needs is not to be cucked in the training stage: if it's censored then it cant do anatomy well.pony is trained on EVERYTHING so it can make hands and figures very well, clothed or not, porn or not.
>>101701028anon what's better as an anime non porn model, autismmix models or animagine?
>>101701059animagine, all pony models are made for porn first in mind
>>101701059animagine is good but autismmix is the best so far, even without loras you can prompt random characters by their booru tag and get great results. and then WITH loras it's even better.I use this extension which is really useful:https://github.com/DominikDoom/a1111-sd-webui-tagcompleteif I type fate for example, i'll see stuff like artoria/nero/etc pop up and it shows the popularity of the tag, so if I use the nero tag i'll get a nero, even without a fate LORA.for example here is a mordred I got with a simple mordred prompt:
>>101701037what model is that?>posting old gens while I wait for my 1060 to struggle through the current one
>>101701079>>101701087super helpful, thx anons
Pony founder is discussing finetuning w BFL
>>101701111Both models are good, the great thing about having multiple models is you can get different styles with one click, just like a Lora.
>>101701054>all sd3 needs is not to be cucked in the training stageit will be. I promise you they have no idea what they are doing. Flux was made by the people who invented Stable Diffusion. The only people left at SAI are retarded middle managers.
tl;dr flux? will it kill cokejeets company?
>>101701160open source dall-e, in terms of functionality: demanding on VRAM but good
>>101701160It can into hands.
it's joever
lmao the prompt comprehension is great
>>101701160considering the brain drain and general retardation from cokejeets company I don't think they can redeem
>>101701160who's cokejeets? emad?
>>101700545>clearly readable hiragana
3060 12GB. Thankfully, the biggest slowdown is caused by the lack of VRAM, increasing resolution has much less impact than I expected.
>deleted image of Hamster it was labeled incorrectly as a Kolors gen wheni t was FLux kek..
>>101701841
>>101701854yeah. mine is 1060 6GB and it works, now 1060 24GB is needed.
>>101700608There's not a technical reason preventing it. There just hasn't been the need for that sort of infrastructure since all of the local image models worth using have been so tiny. Now that we're into the 10B+ range you'll see the tools start adapting to it.
>Flux w/SD3 USDU
improved
does comfyui have merge / combine nodes out of the box yet or will I have to risk downloading malware?
>>101700396Why is there a discord in OP?
>>101702855pretty sure it does
>>101702717NOO THIS IS LITERAL GENOCIDE, THINK OF THE POOR SKIBIDI MIKU ARTISTS
>>101703176bah, wrong image
>>101703176>>101703201still doesn't, sad
>>101703201just realized the select menus weren't highlightedadd node - advanced - model merging
>>101702855I wouldnt download anything linked / mentioned here after what debo tried to pull off when he wanted to catch schizo anon
>>101703199>>101702717this image not being done by some twitter folx cost minority artists 100s of dollars
almost perfection
Is this the tranny thread?
>>101703339this is the verification thread for the doxcord silly
Does flux work on a1111?
>>101703583no, they hate voldy for some reason, did not give him access to preview release, so a1111 bros need to wait
>>101703605>former stability employees hate voldyImagine my shock!
>>101703621why do they have him doebeit?
>FLUX ftw
SKIBIDI MIKU
I came out as trans to my parents today.It did not go well.
>>101703605comfy wasn't given any special access. he just downloaded the model and integrated it. nothing is stopping voldy from doing the same, apart from his crippling video game addiciton
Someone please implement this on comfy/autohttps://github.com/SusungHong/SEG-SDXL
>>101703841>comfy wasn't given any special accesshe was though
>source: my ass
>>101703882>>101702282>we have a partnership with comfy org which, amongst other things, lets us ensure that comfyui and swarmui have day-0 (or day-1 when mcmonkey manages to sleep through the entire release day) support for our models
>>101703908LARP
>>101702717Someone with A. I. D. S. on the internet again.
verifying a sandwich
>>101703851this looks promising
>>101704148
>>101703851Will try on comfy, can't promise auto tho
What is the thread theme?
>>101704355chibis
https://files.catbox.moe/4cwqpo.jpgnearly forgot, today is caturday... je suis expecting beaucoup cats today..
>>101704481
>>101704368
cool, 1024x1024 on 3060 12gbthis is with modules that contain weights, its loading and unloading each module as needed, my ssd is slow which doesnt helpill experiment with loading weights to ram and using set_constants, that method would only need 1 module for each block type too, might be fasteralso with the planned workspace changes for lazy loading/free after use (nearly, minus 1 or 2 for FluxTransformerBlock) every module for a block type could be loaded at once, then the layers would run much faster because it would only need to allocate the workspace, loading the next module is slowing it downsee the second progress bar for a demonstration of the potential speed by loading only 8 of the modules for each block type, about 4x faster than it is atm
>>101704670fp8 weights will help tooalso inb4 anon starts moaning again
>Julien
>>101704670sounds too good to be true. what's the catch?
>>101704730Anon, I'm gonna cooom so hard thinking about those juicy modules and thicc VRAM. Mmm yeah, load those weights into my tight RAM daddy. Fuck my SSD harder with those workspace changes. I wanna feel every inch of those FluxTransformerBlocks inside me. Pound my progress bars with your fp8, you dirty l33t h4x0r. AHHH AHHH I'M ALLOCATING!!!!1! >_<
>>101704670so this and AIT could mean vramlets could access flux without having to consider the heat death of the universe?
>>101704781>53s/it
>>101704773idk none i guess>>101704781yeah, it would need someone to develop the ait comfy node again though>>101704855thats about the same as 3060 users are getting atm with shared memory. also you're ignoring everything about how im going to make it faster. idk why people like you have to moan at everything
>>101704921>moanctft?
PE teacher
>>101705061why so much leg? too much leg
If an argument goes on for more than 5 minutes, neither person has any idea what they are talking about
any good tags for someone who's working or focusd on work? there's no danbooru tags for that sadly.
>>101705132so true>>101705156at work, maybe
>>101705156what do you consider a 'focused' expression though?
>>101705156(subject) working on a (whatever)
>sdg and ldg both using lux
>>101705302ze 'lux
>>101705302flux*
>>101705187>>101705296thanks>>101705249like someone foccused on work, oblivious to things other than it
>>101705302maybe it's time for /ldg/ to come home, it's not really even an 'alternative' like pixart and whatever other shit they were using since flux is where the sd devs wentthe other thread has no reason to exist anymore
>>101705340in two weeks most of you will be doing coomer 1girls from pony
>>101705069Evolutionarily, to run long distances.Currently, to please me.
>>101705328>what do you consider a 'focused'>>like someone foccused on workuhh...>>101705340they didn't split because they're a fundamentally different topic. they split because they wanted their own insular thing.
>>101705516gee anon I wonder
How many people you knowWho can name every serial killer who ever existed in a row
>>101705340they wanted to get away from the constant rulebreaking, avatarfagging, julien & comfy public homosexual ERPing and the thread schizo
>>101705605lets be fair the drama is the best part of /sdg/
>>101705605thankfully we still have you, (and me~)
>>101705340kek, is this the new cope? SD itself is based on research done by these guys beforre they joined stability. are you going to say that everything is SD then?
how to prompt The Matrix style green 1s and 0s raining in the background?I wanna do waifu in leather coat and bikini and shades with that as a background
>>101705756matrix digital rain, or matrix code
>>101705792thanks
>>101705813>1024x1792
>>101705639I hope you're doing well friend
>>101705756"matrix code rain" or just "code rain" is what I've seen it referred to as. but its not 0s and 1s, its complex characters
>>101705883that's fine, thanks anon
>>101705340/sdg/ is the SAI shilling thread and is about stable diffusion onlythe drama and schizos are contained here too/ldg/ welcomes all local models and is very friendly in general
>>101705823the sign of a good model that does exactly what you ask for
modified fingers with mspaint
not sure where the straw came from
>>101706162>Anon thinks parasitic crystalline growth structures is straw, cute.
>>101706212blue board
>>101706162>code rain>its just rain
>>101706289so code rain for background?
>>101704309<3
XL vae still runs in lowvram mode with 8GB VRAM after the commit that changed memory management for flux. is it possible to revert a portable install?
What addons do you guys use for alternate methods of downweighting prompt terms? I used to use BlenderNeko's addon but it hasn't been updated in 4 months and doesn't work with SD3 last I checked
>>101705631Enjoy the dead general
might do a ten thread recap, I haven't done one in ages since I haven't been around
>>101706706I can't wait until you do one of those
>>101706634it now also goes into lowvram mode whenever i use ipadapter, damn, i am going to just paste the 27 build on top of my current instal and hope for the best
>>101706706did you see a waifu diffusion V demo released?
>>101706747waifu diffusion is still a thing? woa didnt hear from them since late 2022/ early 2023
>>101706747No, what's the base? Sounds exciting but can't get distracted rn I have another 1200 images to look through still
>>101706768and its still just as bad too
Morning anons
>>101706880
>>101706880Morning
>>101706722that fixed it, guess i just won't update this install ever again
schnell ain't worth it, has anyone figured out how to prompt styles consistently with flux? i try for something like monet and it just does a generic painting and the facial features end up being photo-like
>>101706926>prompt styles consistently with fluxI've seen people say to use the clip encoder for style input. idk how well it works, but its something you can try out
>>101706880G'mornin Quokkas, Chibis theme
>>101706951i've tried that but it honestly doesn't seem to do a very good job (flux d on the same prompt as above), shit takes 5 minutes to gen tho vv sadi hear flux is still good at small resolutions so i might try that and then upscale with SDXL
>>101706926>Impressionist traditional painting, oil paint on canvas, in the style of Claude Monet. --YOUR PROMPT--set guidance to 1.5-2 run 35 steps
>>101707047Been a while since we've seen a chimeric creature
>>101707007thanks i'll try that, though i think i've done similar in the past i usually put it at the end of the prompt
>>101707087
>>101707121there is some debate on how txxl5 sees prompts, but, yes style as infront helped alot
>>101707137>>101707047can you share prompt?
>>101707245A faded dark fantasy film still captured on a worn-out VHS tape, exhibiting desaturated hues with a subtle blue tinge. The image is characterized by grainy texture, visual noise, and distracting tracking errors causing misaligned video signals, wobbles, and jarring jumps. Blocky compression artifacts, pixelation, and ringing add to the vintage aesthetic, while cable noise and interference introduce snow, fuzz, and ghostly apparitions floating across the screen, enhancing the VHS unique aesthetic. Being a hybrid of different creatures, the Sanctuary Guardian possesses various abilities. Its body is part lion, part goat, and has a long, spiked scorpion tail, in addition to two sets of wings.Its red-eyed lion half has white fur and large paws, which it uses to strike its opponent. Its goat legs have a dark purple color and can perform powerful kicks, and it uses its large set of curved horns to ram its opponent. Its four large wings give it the ability to fly for short periods of time and to release powerful blasts of wind, and its scorpion tail has poisonous properties. In addition to these deadly abilities, the beast can spew lightning projectiles from its mouth.
are we still BREAKing our prompts
>>101707142i wish we could just train one of these modern models to do artist styles as well as SD 1.5 couldthis is the same prompt but with your monet thing in front
ok 1.5cfg FLUX is insane
I'm not even sure how this one happened, but either way I'm here for it>a (centaur) woman outside at a farm, tied shirt, midriff, side slit skirt, wiping sweathttps://litter.catbox.moe/fjomwa.png
>>101707279huh ok, t5 prompting is wild, you gotta just tell the model a whole ass story now?
>>101707343or have a llm do it for you. half of it is a dark souls character description from a wiki btw.
>>101707353yeah if i could fit an LLM *and* Flux into my GPUs T__TI might get a $50k grant to work on AI stuff, then i'll get myself a couple A6000s to make flux waifus
>>101707376ollama with a 7-9B parameter llm doesnt take much to run
>>101706926Kinda shitty that they made the dev version non commercial (so the big finetuners will not touch it). Schnell is WAY worse than dev despite being the same size.
>>101707376What model and prompt anon? She's lovely.
>>101707448flux dev 3.5 guidance, same prompt in clip and t5, 285 seconds to gen>Impressionist traditional painting, oil paint on canvas, in the style of Claude Monet. a close-up half-portrait photo of a gorgeous young hapa (half korean) woman wearing a tight pink satin gown with lace trim, she's wearing dark eye makeup and shimmery lip gloss, she has a knowing seductive smirk on her face, she is thicc with large breasts, she is leaning forward showing ample cleavage and a sparkling necklace hanging low between her breasts, she looks elegant and wears fine jewelry, she is sitting at a bar, lit with colorful neon lights
Can someone explains the difference between SDXL, SD Cascade and SD3?I've been out since sd1.5 and didn't keep up with all the new base models.
>>101707489>Impressionist traditional painting, oil paint on canvas, in the style of Claude MonetNot denying the result is kino, but it's concerning how none of this is reflected in what got generated.
It's over. FLUX won.It feels like a true successor to 1.x stable diffusion.I haven't put it through some of my prompt comprehension tests (where Dall-E failed badly) but its performance on normal photos is amazing.
>>101707511SD3 is garbage. Don't bother with it. The development was a mess. The SD3 that was released to the public is unusable. Cascade was made by another team and released by SAI. It's kino and really fast, but the community didn't really supported it. SDXL is like a better 1.5 for 1k images. At first it was kind of lackluster, but the community has worked a lot on it and by now there are some very good fine tunes of it.
>>101707548that kind of photo stopped being normal in 2003
>>101707489Thanks, I didn't even think flux could make anything outside of uggos or ordinary looking women.
>>101707559Alright, SDXL is it then, I'll see what it can do.
>>101707548it's great at realism and anime but if you want anything in between like an oil painting it's trash, just straight up ignores that part of the prompt esp at smaller sizes (this was supposed to be an impressionist oil painting), also just SO slowi miss 1.5 ability to recognize and emulate artist styles and nothing since has come even remotely close
>>101707548prompt?
>>101707559why does no one talk about cascade anymore, what was it's issues?
I'm generating scenes with +2 people and using pic related to try to detect the people. It's working good and properly detects each person with a proper silhouette. The issue I'm having is that they come out as one entire segment rather than separate segments. Is there anything that can be done to either separate the people into individual segments or just have the detector work on a specified region of the image?
>>101707601Can't say for sure, because I haven't used it myself, but some anon was getting some good painting-like results by changing the sampler. I think they were using Euler Karras.
>>101707617I lost it, but it was something like>My super busty cousin Winnie Liu's 21st birthday at uncle David's flat in San Francisco. 1998. She always had a very big chest. That weekend she was rocking a bustier bodycon midi with tons of cleavage. grainy photo from an old family photo album.
>>101707674Thx anon. T5 prompting is so schizo damn
>>101707674>>101707548would
>>101707740That's what my prompts have looked like since the SD1.4 days. Nothing has changed
im using a 3090 and fp8 and this shit still takes a minute per image
>>10170782045-50s in fp16 for me, but it's on its own dedicated machine.I'll get a second 3090 sometime next week to test if the trick about running vae and clip will speed up stuff.
NGL this model is amazing. I'm running flex dev on a 3090; it IS a little slow but quality >>>> quantity all the way."A still screenshot from a visual novel from the mid 2010's. A cute naked (black:1.1) girl with (massive breasts:1.5) is on the left yelling angrily at a scared Japanese child on the right, with pink cute text on the bottom bar saying "Stop being a little bitch and suck my nigger tits!" "
so /ldg/ won?Sad.
Here's another version, slightly longer:>A still screenshot from a visual novel from the mid 2010's. A cute (black:1.1) girl with (massive breasts:1.5) is on the left yelling angrily with a wild expression at a scared Japanese child on the right, with pink cute text on the bottom bar saying "Stop being a little bitch and suck my nigger tits!". Very high quality, extremely detailed, 8k style of anime like a dedicated still CG. The background is a high school classroom with windows on the left and a beautiful ocean view outside. thin transparent menu bar on top and option b8uttons on right.
For some reason Flux like to put heart pasties on nips, but when nips shows up they look deformed, rarely a good onehttps://files.catbox.moe/so79rl.png
>>101708155I have also noticed this. Even when I'm explicit about the nudity (and often when I'm not and don't even specify it!) it seems to heart pastie just as often as it has normal nipples just as often as it removes nipples entirely! Weird trait, but I can fap to it.
>>101707963
>>101705249>>101704773>actual readable text in /sdg/ on local Neat. Are these first tries or is there inconsistency?
Had to do like five gens, but it got the gist of it.>Jesus working at a depressing booth with papers and pencil in hand, and a line of niggers coming from the foreground to him. He looks exasperated and tired. Above the booth is a sign saying "Make A Wish Foundation" and a small star of david on the right of the sign. In the background there are coca cola and mcdonalds posters with bold black caption "CONSUME." on the left. On the right wall there are Amazon and Google posters with bold RED caption "OBEY".
just got back from my run. mass reply time>>101707296I've been wondering that as well. does flux support conditional operations? I could try, but I prefer the mystery >>101707326that doesnt look fully denoised>>101707376>yeah if i could fit an LLM *and* Flux into my GPUs T__TI have a lot of gpt credits but I'm afraid to plug in an LLM node now....>>101707548>It's over. FLUX won.lots of people are saying it cant be trained. that will determine whether its a long term option or not
>>101707657>>101707601yeah i have problems getting it to do art styles too, i'll try karras
>>101708301I don't understand this argument. How could it possibly not be trainable? How do you think BFL made the model in the first place? Of course it can be trained; people are just butthurt that they have to rent out an A100 instead of making garbo-tier LoRA's on their 980
>>101707963>visual novelsoooo, good idea>>101708155>pasties on nipsdamn if only pasties anon were here >>101708197they're mostly all first tries
>>101708310ok interesting, same prompt same seed, just set to karras, definitely more art style
>>101708355next two seeds just seems kinda noisy and fucked up tho, not really doing the impasto thing like the first one
>>10170837730 steps instead of 20, fml whyyy not it's going to take even longer
>>101707909>>101707963kek. do one with shinji?
>>101707963Prompt please
>>101708338>How could it possibly not be trainable?some people say the liscence, some people say the hardware reqs, some people say the architecture, some people say the available weights. idk how much of it is truth and how much of it is FUD
>>101708431mixed signals from BFL. The license is permissive to finetune, but the devs say it can't be done
>>101708429It's the greentext there. That's the whole prompt. I did fp16, euler, and simple.
>>101708429>im a retard
>>101708431Schnell is open source and allowed to be tuned, so some effort will be spent to reverse the distillation and then someone is going to run on epoch on top of it with 2 million images and voila.
>>10170842340 steps euler karras
>>101708463>>101708467I'm really retarded today, sorry, thanksI don't know how I missed this wall of text omfg
>>101708473I've been out of the space for almost a year now so I'm a little out of the loop. What exactly does BFL mean when they say it's a distillation? It doesn't sound like a pruning since dev and schnell are both 12B (and we don't know how large Pro is). What's going on architecturally?>>101708424I can excuse pedophilia and embrace racism, but I draw the line at mecha anime.>>101708455Do not underestimate 4chan autism
>>101707489>>101708505tried with this prompt and it loses a lot the style, i'm going to ask claude to write a 2 paragraph description of impressionist techniques to see if it balances out
so how exactly do i install flux locally from the hf repo? is there a tldr for dumbretards for where to put each file?
>>101708473simply not happening
>>101708558the same place you put your other models
>>101708567simply will happen, sorryI know your plan was to have a monopoly behind a paywall, but it's happening and you have PHDs that want to fap locally too
>>101708558flux1-dev.sft/flux1-schnell.sft into models\unetae.sft into models\vaeclip_l.safetensors and t5xxl_fp16.safetensors or t5xxl_fp8_e4m3fn.safetensors into models\clip
reminder open source always wins, dall-e wont even let you prompt certain things cause "controversial".
>>101708431It's all of those combined. Biggest obstacle: It was never meant to be finetunable and BFL went to great lengths to ensure that you can't simply take the schnell model, finetune it a bit and spin up a competing firm that offers better product than their Pro model. Everything they have done is to ensure that it will never happen and them releasing it under Apache 2.0 means they are pretty damn confident that someone like NAI can't take their model, train it with anime tits and then start competing with them.
>>101708515You have the Pro model which is what the training happened on. That model becomes a teacher. You then have the Schnell model, it's designed to run fast. Instead of training on 100 steps per image, you do 4 steps on Schnell and the training is strictly comparing the output of the Teacher with the Student. In the process the Student gets crippled and loses a lot of the trainability but it still works for the target goal (fast and quality image generation).
>>101708567What makes you think it wouldn't? The community has been waiting for over a year for an SD 1.5 replacement, never got it, and has been hoarding compute and desire until Flux hit. Only real achievement of note in the interim was Pony.
>>101708607also it can do a wide range of styles, like this: "in the style of a 1940s comic":
>>101708622Read: >>101708609
>>101708609It's hilarious hubris honestly, everything can be reversed. The question is how difficult it will be to get the compute but honestly when you have a very obvious end goal and target it makes research easy because the research simply is>we want to train Schnell on a 24 GB GPUMany people will be taking a crack at it and we already know the power of boners and there are many PHDs who literally have access to H100s provided by their universities.
>>101708600oop i focused too much on the folders while sft was in mainthanks anon
>>101708640We're talking about computers not magic spells.
>>101708633>>101708633>>101708633
>>101708649I think flux has done more harm to open source than good. People are now gonna wait for something that will never come and will not move on.BFL is not and does not want to be a Stability. They have different goals. They are looking into competing with MJ and Dall-E, Sora and any video model MJ eventually makes. Their business model is also not similar to what Meta is doing with llama. These open access models are pure marketing and never meant to be the models open source community adopts as the new base to build upon.
>>101708744>I think flux has done more harm to open source than goodthe greatest harm was showing people what a good local model can do. when SAI releases SD3.1 and it can't compete with flux, a lot of people are gonna write off SAI for good at that point.
>>101708744>>101708744>>101708744Then we, the community, will have to adapt, because what you're arguing -- in essence -- is that instead of waiting for a crowdfunding or enthusiast to spend a couple grand a hundred hours on a pro bono finetuning project, we should instead wait for a nonexistent new tech startup to spend a couple hundred million and tens of thousands of PhD man hours training a new model from scratch, pro bono. What is with this fatalistic "full open source or bust" nonsense? SAI is not coming back to save us, and if that's the case we might as well adapt to the new environment and be grateful for the amazing models we're given, and collectively organize to pool our efforts into fewer high quality fine-tunes instead of many, low-quality ones.
>>101708609I find it difficult to believe "competitive" means anything less than attempting to train a full foundational model. We have the weights. Time will tell.
>>101708864>1708744>>I think flux has done more harm to open source than goodSkill issue.
>>101708338Dev and schnell are distilled models. They were basically trained to imitate Pro. They can only be trained with access to the Pro weights, because Pro acts as a "teacher" model. See this github thread: https://github.com/black-forest-labs/flux/issues/9
>>101709169>https://github.com/black-forest-labs/flux/issues/9See https://x.com/ostrisai/status/1819802556261863925?t=NaBwMCu3sz4_fibQ8p2ATQ&s=19
>>101707928Who fucking cares who "won"?
