Research map / Video / Video generation
Adaptive Caching for Faster Video Generation with Diffusion Transformers
Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S. Ryoo, Tian Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add AdaCache-fast (adaptive caching) to Open-Sora | Open-Sora 480p · Open-Sora 480p-2s | speedup: 1.00x -> 2.24x | Improved |
| use AdaCache-fast instead of PAB-fast | Open-Sora-Plan · Open-Sora-Plan 512x512 | speedup: 1.45x -> 3.70x | Improved |
| use AdaCache-fast instead of PAB-fast | Latte · Latte 512x512 | speedup: 1.33x -> 2.74x | Improved |
| add motion regularization to AdaCache-fast (ablation) | AdaCache-fast · Open-Sora-Plan 512x512 | VBench (%): 75.83 -> 79.30 | Improved |
| add motion regularization to AdaCache-fast (ablation) | AdaCache-fast · Open-Sora 480p | VBench (%): 79.39 -> 79.48 | Improved |
| add motion regularization to AdaCache (ablation) | AdaCache · Open-Sora 720p-2s | VBench (%): 83.40 -> 83.50 | Improved |
| add motion-gradient as early motion predictor (ablation) | AdaCache + MoReg · Open-Sora 720p-2s | VBench (%): 83.36 -> 83.50 | Improved |
| estimate motion over multiple steps of frame differences (ablation) | AdaCache + MoReg · Open-Sora 720p-2s | latency (s): 93.50 -> 95.65 | Hurt |
| use cosine distance instead of L1 for the cache metric (ablation) | AdaCache · Open-Sora 720p-2s | VBench (%): 83.40 -> 83.19 | Hurt |
| compute cache metric at start instead of mid-DiT (ablation) | AdaCache · Open-Sora 720p-2s | VBench (%): 83.40 -> 83.30 | No clear effect |
| use MLP residual instead of temporal attention for cache metric (ablation) | AdaCache · Open-Sora 720p-2s | latency (s): 89.53 -> 99.72 | Hurt |
| use slow basis cache-rates instead of fast (ablation) | AdaCache · Open-Sora 720p-2s | VBench (%): 83.40 -> 84.12 | Improved |
Closest papers by meaning
- LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation (NeurIPS 2025 · similarity 0.88)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.86)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.85)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.85)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.85)
- VideoScene: Distilling Video Diffusion Model to Generate 3D Scenes in One Step (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
