Software Engineering Courses (SECourses)•15mo ago

saw an interesting comment. are different blocks responsible for face likeness and overall image loo

saw an interesting comment. are different blocks responsible for face likeness and overall image look? or its just broscience?

Ddaroth saw an interesting comment. are different blocks responsible for face likeness a...

crystalwizard•10/5/24, 2:54 PM

Each block does something specific

EEreshkigal Click to see attachment

Furkan Gözükara SECourses•10/5/24, 5:19 PM

yes more memory better. i plan to upgrade 128gb

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

Furkan Gözükara SECourses•10/5/24, 5:20 PM

true but i am planning models to get more ready

Furkan Gözükara SECourses•10/5/24, 5:20 PM

still too early i think

FFurkan Gözükara SECourses yes more memory better. i plan to upgrade 128gb

crystalwizard•10/5/24, 5:21 PM

you need vram, not system ram

Furkan Gözükara SECourses•10/5/24, 5:22 PM

ok i see his method seems very weak

Furkan Gözükara SECourses•10/5/24, 5:22 PM

https://huggingface.co/ostris/OpenFLUX.1/discussions/6#67009c9d23df5a7bd8628f4a

ostris/OpenFLUX.1 · how did you train it?

Furkan Gözükara SECourses•10/5/24, 5:22 PM

no i doubt that it will ever work

Ccrystalwizard you need vram, not system ram

Furkan Gözükara SECourses•10/5/24, 5:22 PM

i need both :d

FFurkan Gözükara SECourses i need both :d

crystalwizard•10/5/24, 5:25 PM

not really. you shouldn't be using that CPU for anything graphics related. and if you have an AMD, you should go through CS1o's AMD guides and make sure it's set up right.

crystalwizard•10/5/24, 5:25 PM

everything should be on your GPU and in vram

Ddaroth saw an interesting comment. are different blocks responsible for face likeness a...

Furkan Gözükara SECourses•10/5/24, 5:26 PM

it is pretty much nonsense

Furkan Gözükara SECourses•10/5/24, 5:26 PM

actually i was expecting training single block would reduce bleeding but it didnt

Furkan Gözükara SECourses•10/5/24, 5:27 PM

i guess due to transformer arcihecture even single layer changes alls output

Ccrystalwizard not really. you shouldn't be using that CPU for anything graphics related. and i...

Furkan Gözükara SECourses•10/5/24, 5:27 PM

it depends on how many apps you open at the same time

FFurkan Gözükara SECourses it depends on how many apps you open at the same time

crystalwizard•10/5/24, 5:31 PM

there's a 12 step program for that

xlads•10/5/24, 6:29 PM

@Dr. Furkan Gözükara have you tried flux 1.1?

Xxlads @Dr. Furkan Gözükara have you tried flux 1.1?

Furkan Gözükara SECourses•10/5/24, 6:29 PM

it is only api

Furkan Gözükara SECourses•10/5/24, 6:29 PM

i did a quick test it was better for some mystical stuff like

Furkan Gözükara SECourses•10/5/24, 6:29 PM

a dinasaur

Furkan Gözükara SECourses•10/5/24, 6:46 PM

backing up test data

Furkan Gözükara SECourses•10/5/24, 7:45 PM

trainings started

Furkan Gözükara SECourses•10/5/24, 7:45 PM

dxqb•10/5/24, 7:53 PM

I'm not comfortable with you spending 8 GPUs on my 5 lines of code from thing morning

dxqb•10/5/24, 7:53 PM

but it's up to you

Furkan Gözükara SECourses•10/5/24, 8:09 PM

ye it is fine

Furkan Gözükara SECourses•10/5/24, 8:09 PM

if works it will be big

FFurkan Gözükara SECourses if works it will be big

Maxivi•10/5/24, 8:30 PM

are you using your 5200 reg images?

Maxivi•10/5/24, 8:30 PM

i dont think dx is using any reg images

MMaxivi are you using your 5200 reg images?

Furkan Gözükara SECourses•10/5/24, 8:34 PM

yep

Furkan Gözükara SECourses•10/5/24, 8:34 PM

i am testing both cases

dsienra•10/5/24, 9:11 PM

https://huggingface.co/nyanko7/flux-dev-de-distill I'm testing this model on forge and I'm having interesting results, distilled cfg does nothing, is completely disabled

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

dsienra•10/5/24, 9:12 PM

regular Flux-dev random prompt

dsienra•10/5/24, 9:17 PM

de-distilled model same prompt same seed

dsienra•10/5/24, 9:17 PM

on the de-distilled model if you prompt random you get random noise

dsienra•10/5/24, 9:18 PM

on the normal flux-dev you get a coherent output because is allways guided

Ddsienra de-distilled model same prompt same seed

Furkan Gözükara SECourses•10/5/24, 9:21 PM

ye probably these are just failed experiments

Furkan Gözükara SECourses•10/5/24, 9:21 PM

not true destilation

dsienra•10/5/24, 9:22 PM

If your prompt is not random you get very similar outputs on both models

dsienra•10/5/24, 9:22 PM

de-distilled model

Ddsienra de-distilled model

Furkan Gözükara SECourses•10/5/24, 9:25 PM

but with cfg right?

dsienra•10/5/24, 9:26 PM

regular flux dev same prompt

FFurkan Gözükara SECourses but with cfg right?

dsienra•10/5/24, 9:26 PM

yes cfg 3.5

dsienra•10/5/24, 9:28 PM

the results are promising it behaves like a non distilled model

dsienra•10/5/24, 9:29 PM

with cfg at 1

dsienra•10/5/24, 9:30 PM

his explanation make sense, idk

dsienra•10/5/24, 9:30 PM

Flux-dev-de-distill
This is an experiment to de-distill guidance from flux.1-dev. We removed the original distilled guidance and make true classifier-free guidance reworks.

Model Details
Following Algorithm 1 in On Distillation of Guided Diffusion Models, we attempted to reverse the distillation process by re-matching guidance scale w. we introduce a student model x(zt) to match the output of the teacher at any time-step t ∈ [0, 1] and any guidance scale w ∈ [1, 4]. We initialize the student model with parameters from the teacher model except for the parameters related to w-embedding.

Since this model uses true CFG instead of distilled CFG, it is not compatible with diffusers pipeline. Please use inference script or manually add guidance in the iteration loop.

Train: 150K Unsplash images, 1024px square, 6k steps with global batch size 32, frozen teacher model, approx 12 hours due to limited compute.

dsienra•10/5/24, 9:31 PM

@Dr. Furkan Gözükara I tried to train it on kohya but it not supported

Ddsienra Flux-dev-de-distill This is an experiment to de-distill guidance from flux.1-dev...

Furkan Gözükara SECourses•10/5/24, 9:33 PM

150k not bad

Furkan Gözükara SECourses•10/5/24, 9:33 PM

12 hours really low though

saw an interesting comment. are different blocks responsible for face likeness and overall image loo

Similar Threads

Similar Threads

Similar Threads