Efficient Video Diffusion Models: Advancements and Challenges

Shitong Shao, Lichen Bai, Pengfei Wan, James Kwok, Zeke Xie

TL;DR

This repository accompanies the survey Efficient Video Diffusion Models: Advancements and Challenges.

The survey studies how to make video diffusion models practical under severe inference-time compute and memory constraints. It organizes the literature into four main acceleration paradigms: step distillation, efficient attention, model compression, and cache / trajectory optimization.

Beyond taxonomy, the survey focuses on two deployment-oriented questions:

• how existing methods reduce the number of function evaluations
• how they reduce per-step overhead in spatial-temporal generation pipelines

It also highlights open problems around quality preservation under composite acceleration, hardware-software co-design, robust real-time long-horizon generation, and standardized evaluation infrastructure.

What This Repository Provides

• index.html: a shareable interactive paper atlas organized by the survey taxonomy
• generate_paper_atlas.py: a local generator for rebuilding the atlas from the bundled metadata snapshot
• data/: the frozen input snapshot used by the atlas generator
• link_overrides.json: manual link corrections used during generation
• missing_links.json: generated missing-link report

Quick Start

• Open index.html directly in a browser
• Or serve the folder locally with python3 -m http.server 8765
• Rebuild the atlas with python3 generate_paper_atlas.py

Citation

If you find this survey useful, please cite:

@article{shao2026efficient,
  title   = {Efficient Video Diffusion Models: Advancements and Challenges},
  author  = {Shao, Shitong and Bai, Lichen and Wan, Pengfei and Kwok, James and Xie, Zeke},
  journal = {arXiv preprint arXiv:2604.15911},
  year    = {2026}
}