A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models

Introduction

Welcome to the implementation of our ICML 2024 accepted paper, "A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models." This repository contains the code and resources necessary to reproduce the results presented in the paper. Our framework introduces a novel early exiting strategy to significantly accelerate the sampling process in diffusion models without compromising the quality of generated samples.

Update News

• DiT experiments updated
• UViT experiments update

References

This implementation is inspired by and references the following GitHub repositories:

• DiT : Original implementation of Scalable Diffusion Models with Transformers (DiT)

• U-ViT : All are Worth Words: A ViT Backbone for Diffusion Models (U-ViT)

• Diffusion Models by OpenAI: Implementation of guided diffusion models.

We would like to express our gratitude to the authors of these repositories for their valuable contributions to the field.