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.
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)