I understand what you say, may be you right and is imposible to reverse engineering it
I 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
Cthat depends on what you mean by work, and whether what ostris wants is what they actualy do or not
Dlet's see as I said I'm not an expert, I can't gave a informed opinion
Dyou have good points
Ci've got close to 3500 hours in on studying stable diffusion in the last 2.5 years. and close to 300 hours in on studying flux since it released. if not more
Ci literally spend an entire month sitting here, feeding it one single prompt at a time to see what it would return by default - walking through it's latent space
Cgive it an unclear concept - you're going to get women, dogs, fantasy, or anime cat girls - and frequently a mixture of those
Cnot random, all over the place, images - like you should
Cit's massively overfit, and it's been heavily doctored to deal with the 'woman laying in the grass' warping and shortening issue it shares with sd3-2b-medium.
Dyes that is because of the distilled guidance right?
Cand that makes it impossible to do anything with other than what it is
Cno. that just makes it worse.
Cno, that's becuase of other things that were done before that
Dyou always get something coherent
Csure, because it uses the t5xxl encoder nd the clip_l encoder. so you're going to get coherant. but give it a word like Umber and you should get mostly the color. the dictionary defintion for umber is "a natural pigment resembling but darker than ocher, normally dark yellowish-brown in color ( raw umber ) or dark brown when roasted ( burnt umber ).
2.
a brownish-gray moth with coloring that resembles tree bark."
2.
a brownish-gray moth with coloring that resembles tree bark."
Cthat's not what you get
Cyou should get random subjects, since umber isn't a subject - all over the map
Cyou don't get that
Cyou get very specific subjects as that's what it draws when you don't tell it a subject
Dyou put something random and and get a dog for example
DI know what you said
Cif i tell you this word "Unboundedness" - what would you tell me it looks like?
Dhehe
Dno clue
Dthe model probably returns a cute anime cat
Dhehe
Cit should return random images with random subjects. because it doesn't have a subject so it should just be pulling up data from all over latent space
Cif you put that word into sd 1.5, or sdxl, or midjourney - you'll get interesting results
Cput it in flux dev and see what you get
Dthanks for your detailed explanation, let's see how this ends up
Cbased on the blowup that he had a few weeks back, he'll refuse to admit he cna't do what he wants. he'll continue on till he gets 'something' and then boast about it
Cand it'll return some sorts of results, and people will believe he undistilled flux because they don't know any better. and blackforest will sit there and laugh at him
DI know that is imposible to convert a distilled model to its original state, the information that was lost in the process is lost forever
Cthat's true too. you spend the money, you no longer have the money
Cwe get requests along those lines for Photoshop all the time "hi, i have this photo of this car, can you make it so i can read the license' - and the plate is just a black and white blur. 'um, no. this isn't camera raw footage, it's a jpg"
Dwhat could be a success is to make a model that is more flexible and more trainable, will not be the black forest pro model, will be other model that is more trainable
Dthat if its possible
Cto do that, robin will need to scrap flux and start over. and that was at least a million, possibly more, to train. that's a huge hit now that he's nto at stability.ai