A field guide to faster diffusion transformers

What it does

This repository is a curated index of research papers focused on making diffusion model inference faster and cheaper. It organizes work across sampling strategies, caching, parallelism, quantization, and attention into sortable tables, linking each entry to its arXiv PDF and to implementation code where it exists. Think of it as a field guide to the many tricks researchers have found to make image and video generation less computationally expensive.

The interesting bit

The list explicitly distinguishes between UNet-based and DiT-based caching approaches with architecture diagrams, and it flags which papers have publicly available code versus which remain paper-only. Coverage spans the full diffusion ecosystem, from Stable Diffusion and SDXL to video models like Sora, CogVideoX, and WAN.

Key highlights

• Organizes papers across five optimization categories: sampling, caching, parallelism, quantization, and attention
• Tags each entry with a recommendation level (⭐️ to ⭐️⭐️) and flags missing code with a ⚠️ symbol
• Includes visual diagrams comparing UNet-based and DiT-based caching strategies
• Covers both image and video diffusion, including recent video-specific caches like AdaCache and TeaCache
• Maintains a companion list for LLM inference acceleration

Caveats

• Several listed papers are marked with ⚠️, indicating no linked code repository is currently available
• As a bibliography rather than a framework, it does not implement or benchmark the methods itself

Verdict

Useful as a roadmap if you’re building diffusion inference pipelines and want a quick survey of the acceleration literature. Not what you need if you’re looking for a single library to drop into your project.