This repository curates research papers and code for accelerating diffusion model inference, focusing on techniques like sampling, caching, parallelism, and quantization. It serves researchers and engineers seeking to optimize the performance of diffusion models, particularly for image and video generation tasks. The primary benefit is a centralized, categorized resource for exploring state-of-the-art inference acceleration methods.

How It Works

The project functions as a curated list, categorizing papers by their core inference acceleration technique. Each entry includes a title, publication date, link to the paper (PDF), and a link to the associated code repository where available. The organization into sections like Sampling, Caching, Parallelism, Quantization, and Attention allows users to quickly navigate and find relevant research.

Quick Start & Requirements

This repository is a curated list and does not have direct installation or execution commands. Users are expected to follow the provided links to individual research papers and their associated code repositories for setup and usage.

Highlighted Details

• Comprehensive coverage of diffusion inference acceleration techniques.
• Extensive list of recent (2023-2025) papers with code links.
• Categorization includes Sampling, Caching, Parallelism, Quantization, and Attention.
• Recent addition of CacheDiT toolbox for Diffusion Transformers.

Maintenance & Community

The repository is maintained by xlite-dev. Contributions are welcomed via pull requests and starring the repository.

Licensing & Compatibility

The repository is licensed under the GNU General Public License v3.0 (GPL-3.0). This license is copyleft, meaning derivative works must also be licensed under GPL-3.0. Compatibility with closed-source projects may be restricted due to the copyleft nature of the license.

Limitations & Caveats

This is a curated list of external resources; the repository itself does not provide implementations. Code availability is indicated by links, but not all papers have publicly available code. The "Recom" column uses a simple star rating, which is subjective.