what I want is to train the models without bleeding and concept collapsing
what I want is to train the models without bleeding and concept collapsing
Ci just spend a month walking through the flux dev and flux pro latent space, one token at a time. what was done to flux to hide the warping and shortening issues that everyone saw with the sd3-2b-medium release (because flux is the exact same neural network and setup) is really ugly, and undoable
Cthen train SDXL - flux is also untrainable.
Cyou'll get 'something' but you won't get a real flux train
Dyes Im doing that, and flux for some things
Cflux is extremely rigid as well. and breaks easily
DI trained flux with some big datasets off 3500 images for 40 epochs and I hade great results but is for a style and some new concepts, is unable to train many specific individuals consistently
Cyou'll get 'something' - but you won't get something that actually updates flux's weights the way a training is supposed to. all you're going to get is something that overwrites what the AI knows. the results you'd get would be different if it was actually training the model
Cthat's what a LoRA is for - to allow you to add information and update the weights for specific things without retraining the entire model. but what you're getting is about the same as if someone took a piece of paper, drew on it, and pasted it over a photo. you're not changing the photo, you're just putting somethign over it.
Cand because flux IS NOT trainable, regardless of what you keep seeing people say, when you try to train it for things that do require those weights to be updated, it can't happen
Cyou can make something that'll put your face on an image - but you could do that with face swap, too
Dno ,but face swap not compares at all one you can get training a good lora
Dyou can get a good style
Dand then a lora for one subject at a time
Cand that is exactly what you will not get out of flux. use SDXL
Calso, i'll be very surprised if black forest releases any updates to dev or schnell anytime soon. the money is in pro, and the people that are paying 5 cents a generation for it's API
DI'm not interested in paying for it
Cthat's the only way you're going to get your hands on pro. it's only available via an expensive API
DI use everything local in my computer
Cas do i
Cand pro is the only version that wasn't run through the distill process
Dyes I know
Cgive flux some tokens with unclear subjects and see what it creates for you
Cactually, pull up flux dev and give it the word Pail as a prompt. just one word. run that 10 times, and see what it does
Cuse the actual flux dev that bfl released, not something that's been messed with by the community
Dlet's see, for know it works for my purpose, I can train a style and some concepts and on top a character at a time, that works perfect for me, is tedious because I need a lora for each subject but is a workaround
Dtwo people in the same lora is imposible
Cno it's not. well, for flux it is, but no it's not
Dwith sdxl I can train more than 20 people on the same lora no bleeding and the style at the same time, works perfect
Dbut the quality is so much worse than blux
Dbut the quality is so much worse than Flux
Cbecause flux is not trainable. you are getting 'something' - but you are not getting an actual object trained on flux.
Cyou're just getting something that'll draw over it and run with it without errors
Cthere's a reason that BFL never released the training parmeters - it's not trainable
Cand they know it isn't
Cthey couldn't fix the issues it has, so they did stuff to mask them, and that makes it almost unusuable for anything other than exactly what they released, with a very narrow range of what it can successfully do
Cyou could take sd3-2b-medium, train loras for it, get better trained loras and have stuff that would work just as well.
DI understand what you say, may be you right and is imposible to reverse engineering it
Ci know what they did - along with filling up their data set with midjourney images.
MI haven't spoken much about my ongoing project to de-distill schnell to make a permissive licensed version of flux, but I have been updating it periodically as it trains. I just noticed it is the #2 trending text-to-image model on Hugging Face. Working on aesthetic tuning now.
10/5/24, 2:18 AM
Cit's not a matter of reverse engineering it, it's a matter of it's being impossible to undo
Cyeah, i know. he can't do what he think he can.
Cit's sort of like this - if you mix up flour, water, eggs, chocolate chips, sugar, baking soda, salt - you have a batter that can be turned into a lot of things. YOu could make a chocolate chip cake or cookies at this point. if you then bake cookies, you have cookies. and no one can come along and turn those cookies back into batter.
Cthey can soak them in water, and get mush
Cbut it's still nto batter
Cyou can then shape the mush into some other shape, and say you turned it back into batter and now you've baked a cake or something - but you didn't actually do that. you just made mush and reshaped cookies into cake shaped cookies that now don't taste all that good
Dif you right will see all this projects fail
Mi mean i think he's right, i tried 100+ trainings on these distill models, it just gives it texture and artifacts
Ceh, he'll keep working on it till he gets something that he'll then prance around and boast about. you'd think HE would know better - but he's backed himself into a corner
Che's desperate to be able to create controlnet for flux - and you can't do that, either.
M
