This repository is a curated collection of research papers and resources focused on accelerating generative AI models, particularly diffusion models. It targets researchers and engineers working on improving the efficiency of text-to-image, text-to-video, and other generative tasks, offering a comprehensive overview of techniques like fast sampling, pruning, quantization, and distillation.

How It Works

The project acts as a comprehensive index, categorizing and linking to academic papers that propose novel methods for generation acceleration. It covers a wide array of techniques, including optimizing sampling schedules, reducing model size through pruning and quantization, knowledge distillation, efficient attention mechanisms, and deployment optimizations. The organization by technique allows users to quickly find relevant research for specific acceleration challenges.

Quick Start & Requirements

This repository is a collection of research papers and does not have a direct installation or execution command. Users will need to access the linked papers and their associated code repositories for practical implementation.

Highlighted Details

• Extensive categorization of acceleration techniques including Fast Sampling, Pruning, Quantization, Distillation, Cache Mechanism, Efficient Attention, Dynamic Neural Networks, and Deployment Optimization.
• Features recent research, with many papers from 2024 and accepted for conferences in 2025.
• Highlights specific contributions like "ToCa" achieving significant acceleration on DiT models.
• Includes links to papers and code for each listed technique.

Maintenance & Community

The repository is actively maintained, with recent updates and news regarding accepted papers and new related repositories. Contributions are welcomed via email.

Licensing & Compatibility

The repository itself is not software with a license. The licensing of individual papers and their associated code would need to be checked on a per-project basis.

Limitations & Caveats

This is a curated list of research papers and not a runnable software library. Users must independently find, evaluate, and integrate the code from the linked sources, which may have varying levels of maturity, documentation, and dependencies.