Awesome-DiT-Inference
A curated list of research papers and code for diffusion-model inference, covering sampling, caching, quantization, parallelism, and attention optimizations for image and video diffusion transformers.
Awesome-DiT-Inference is a curated list of papers with accompanying code on making diffusion-model inference faster, covering sampling methods such as DDIM and DPM-Solver, caching techniques, quantization, parallelism strategies, and attention optimizations for diffusion transformer models used in image and video generation, such as Stable Diffusion, Flux, and Sora-like video models.
Each entry links a paper and, where available, its code, and the list is organized by technique rather than by model. It is a companion to a separate list covering large-language-model inference.
The project also tracks a related tool, Cache-DiT, adding features such as ring attention with batched point-to-point communication and hybrid parallelism strategies. It is released under the GPL-3.0 license.