Diffusion Inference Acceleration
Methods for accelerating diffusion model inference at test time without retraining, including caching, pruning, and distillation techniques.
6 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026
Image generators can outperform their multi-step originals on image statistics
Uses adversarial training to compress multi-step text-to-image diffusion into a one-step generator that matches or beats the original on image statistics.
released code
Oct 2026 · InfRec, The Chinese University of Hong Kong, Shenzhen
Video generators can run twice as fast without retraining
Proposes training-free caching and pruning to accelerate video diffusion transformers by nearly 2x.
2 further papers
Nov 2026 · Shanghai Jiao Tong University, Xi’an Jiaotong University
Streaming video generators can now output frames at near-watchable speed
Staggers denoising and cache-refresh stages to achieve real-time video generation rates, accelerating diffusion-based inference.
Nov 2026 · Shanghai Innovation Institute, Fudan University
High-resolution image editing gets a 4x speedup by keeping a low-resolution anchor
Proposes FLASH for 4x speedup in high-resolution image editing using low-resolution anchor and sparse high-resolution branch.
4 of 6 papers shown