flux dev de distill

Introduction

The FLUX-DEV-DE-DISTILL is an experimental model designed to explore de-distillation of guidance from the FLUX.1-dev model. The process involves removing the original distilled guidance to implement true classifier-free guidance.

Architecture

This model reworks the guidance scale using a student model x(zt) to match the output of the teacher model at any time-step t ∈ [0, 1] and any guidance scale w ∈ [1, 4]. The student model is initialized with parameters from the teacher model, excluding those related to w-embedding. This approach adheres to Algorithm 1 from the paper "On Distillation of Guided Diffusion Models."

Training

The training of this model involved 150K Unsplash images sized at 1024px square. The process consisted of 6,000 steps with a global batch size of 32, using a frozen teacher model. Due to limited computational resources, training took approximately 12 hours.

Guide: Running Locally

1. Clone the Repository: Use Git to clone the repository containing the model and scripts.
2. Set Up the Environment: Ensure that you have the necessary dependencies installed, typically using Python and a virtual environment.
3. Run the Inference Script: Since the model is not compatible with diffusers pipeline, use the provided inference script to perform tasks.
4. Cloud GPU Recommendation: For optimal performance, especially if lacking local GPU resources, consider using cloud services like AWS EC2, Google Cloud, or Azure for GPU access.

License

The FLUX-DEV-DE-DISTILL model is licensed under the MIT License, which allows for flexibility in usage and modification.