Software Engineering Courses (SECourses)•15mo ago

@Dr. Furkan Gözükara Hello dr, I have downloaded today two claimed de-distilled flux models, I didn'

@Dr. Furkan Gözükara Hello dr, I have downloaded today two claimed de-distilled flux models, I didn't hade time to testing them, I don't know if kohya_ss is compatible con one of them, one of them is based on flux1-dev and the other in schnell, in one hand the distillation process as I have read is easier to remove on flux1-dev and it have better quality, in the other hand flux-schnell is harder to remove less quality but is apache 2.0 license. you are planning to test some of this models soon?

Ddsienra @Dr. Furkan Gözükara Hello dr, I have downloaded today two claimed de-distilled ...

crystalwizard•10/5/24, 5:08 AM

they can't be undistilled. i've seen that claim posted as well. what he's doing isn't undistilling them, he's jsut breakign them worse then they already are

dsienraOP•10/5/24, 5:10 AM

https://huggingface.co/ostris/OpenFLUX.1

https://huggingface.co/nyanko7/flux-dev-de-distill

this two projects looks promising and legit

ostris/OpenFLUX.1 · Hugging Face

nyanko7/flux-dev-de-distill · Hugging Face

crystalwizard•10/5/24, 5:10 AM

i'm very aware of his projects. he is not undistilling anything

crystalwizard•10/5/24, 5:10 AM

he jsut thinks he is

crystalwizard•10/5/24, 5:11 AM

what was done to flux can't be undone without getting the actual base model and retraining it

crystalwizard•10/5/24, 5:11 AM

and he doesn't have access to that, or the funds to do that, or the machines needed to do that

Ccrystalwizard i'm very aware of his projects. he is not undistilling anything

dsienraOP•10/5/24, 5:15 AM

ok.. may be you right, I'm not an expert or a software engineer

crystalwizard•10/5/24, 5:16 AM

that doesn't mean you might not get good images with his code. you might - but it's not because he's undistilling it

dsienraOP•10/5/24, 5:17 AM

something is doing, dome people claimed that the modified models can produce images without distilled cfg, and they work with a regular cfg scale

dsienraOP•10/5/24, 5:18 AM

what I want is to train the models without bleeding and concept collapsing

crystalwizard•10/5/24, 5:18 AM

i 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

Ddsienra what I want is to train the models without bleeding and concept collapsing

crystalwizard•10/5/24, 5:18 AM

then train SDXL - flux is also untrainable.

crystalwizard•10/5/24, 5:19 AM

you'll get 'something' but you won't get a real flux train

dsienraOP•10/5/24, 5:19 AM

yes Im doing that, and flux for some things

crystalwizard•10/5/24, 5:20 AM

flux is extremely rigid as well. and breaks easily

dsienraOP•10/5/24, 5:23 AM

I 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

Ddsienra I trained flux with some big datasets off 3500 images for 40 epochs and I hade g...

crystalwizard•10/5/24, 5:25 AM

you'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

crystalwizard•10/5/24, 5:26 AM

that'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.

crystalwizard•10/5/24, 5:26 AM

and 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

crystalwizard•10/5/24, 5:28 AM

you can make something that'll put your face on an image - but you could do that with face swap, too

dsienraOP•10/5/24, 5:29 AM

no ,but face swap not compares at all one you can get training a good lora

dsienraOP•10/5/24, 5:30 AM

you can get a good style

dsienraOP•10/5/24, 5:30 AM

and then a lora for one subject at a time

crystalwizard•10/5/24, 5:31 AM

and that is exactly what you will not get out of flux. use SDXL

crystalwizard•10/5/24, 5:31 AM

also, 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

dsienraOP•10/5/24, 5:32 AM

I'm not interested in paying for it

crystalwizard•10/5/24, 5:33 AM

that's the only way you're going to get your hands on pro. it's only available via an expensive API

dsienraOP•10/5/24, 5:33 AM

I use everything local in my computer

crystalwizard•10/5/24, 5:33 AM

as do i

crystalwizard•10/5/24, 5:33 AM

and pro is the only version that wasn't run through the distill process

dsienraOP•10/5/24, 5:33 AM

yes I know

crystalwizard•10/5/24, 5:34 AM

give flux some tokens with unclear subjects and see what it creates for you

crystalwizard•10/5/24, 5:35 AM

actually, 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

crystalwizard•10/5/24, 5:36 AM

use the actual flux dev that bfl released, not something that's been messed with by the community

dsienraOP•10/5/24, 5:37 AM

let'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

dsienraOP•10/5/24, 5:38 AM

two people in the same lora is imposible

Ddsienra two people in the same lora is imposible

crystalwizard•10/5/24, 5:39 AM

no it's not. well, for flux it is, but no it's not

dsienraOP•10/5/24, 5:43 AM

with sdxl I can train more than 20 people on the same lora no bleeding and the style at the same time, works perfect

dsienraOP•10/5/24, 5:44 AM

but the quality is so much worse than blux

dsienraOP•10/5/24, 5:44 AM

but the quality is so much worse than Flux

Ddsienra with sdxl I can train more than 20 people on the same lora no bleeding and the s...

crystalwizard•10/5/24, 5:45 AM

because flux is not trainable. you are getting 'something' - but you are not getting an actual object trained on flux.

crystalwizard•10/5/24, 5:45 AM

you're just getting something that'll draw over it and run with it without errors

crystalwizard•10/5/24, 5:45 AM

there's a reason that BFL never released the training parmeters - it's not trainable

crystalwizard•10/5/24, 5:46 AM

and they know it isn't

crystalwizard•10/5/24, 5:46 AM

they 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

crystalwizard•10/5/24, 5:48 AM

you could take sd3-2b-medium, train loras for it, get better trained loras and have stuff that would work just as well.

dsienraOP•10/5/24, 5:50 AM

I understand what you say, may be you right and is imposible to reverse engineering it

Ddsienra I understand what you say, may be you right and is imposible to reverse enginee...

crystalwizard•10/5/24, 5:50 AM

i know what they did - along with filling up their data set with midjourney images.

Maxivi•10/5/24, 5:51 AM

https://x.com/ostrisai/status/1842388844970135932

hes still working on it

Ostris (@ostrisai) on X

I 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.

Twitter

•

10/5/24, 2:18 AM

crystalwizard•10/5/24, 5:51 AM

it's not a matter of reverse engineering it, it's a matter of it's being impossible to undo

@Dr. Furkan Gözükara Hello dr, I have downloaded today two claimed de-distilled flux models, I didn'

Similar Threads

Similar Threads

Similar Threads