Research map / Video / Video generation
Model Reveals What to Cache: Profiling-Based Feature Reuse for Video Diffusion Models
Xuran Ma, Yexin Liu, Yaofu Liu, Xianfeng Wu, Mingzhe Zheng, Zihao Wang, Ser-Nam Lim, Harry Yang
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 |
|---|---|---|---|
| reuse background-focused blocks instead of foreground-focused blocks (ablation) | ProfilingDiT · HunyuanVideo 720P | LPIPS: 0.1449 -> 0.1203 | Improved |
| reuse background-focused blocks instead of foreground-focused blocks (ablation) | ProfilingDiT · HunyuanVideo 720P | PSNR: 24.23 -> 26.44 | Improved |
| reuse background-focused blocks instead of foreground-focused blocks (ablation) | ProfilingDiT · HunyuanVideo 720P | speedup: 1.57x -> 1.87x | Improved |
| reuse background-focused blocks instead of foreground-focused blocks (ablation) | ProfilingDiT · HunyuanVideo 720P | FID: 58.53 -> 41.10 | Improved |
| use background-only reuse instead of Split background+foreground reuse (ablation) | ProfilingDiT · HunyuanVideo 720P | LPIPS: 0.1753 -> 0.1203 | Improved |
| use background-only reuse instead of Alternate background+foreground reuse (ablation) | ProfilingDiT · HunyuanVideo 720P | LPIPS: 0.1875 -> 0.1203 | Improved |
| use decreasing (Stepwise) reuse intervals instead of increasing (Step Inverse) (ablation) | ProfilingDiT background reuse · HunyuanVideo 720P | LPIPS: 0.2529 -> 0.1203 | Improved |
| use decreasing (Stepwise) reuse intervals instead of increasing (Step Inverse) (ablation) | ProfilingDiT background reuse · HunyuanVideo 720P | PSNR: 24.49 -> 26.44 | Improved |
| use decreasing (Stepwise) reuse intervals instead of fixed (Step Average) (ablation) | ProfilingDiT background reuse · HunyuanVideo 720P | LPIPS: 0.1571 -> 0.1203 | Improved |
| use decreasing (Stepwise) reuse intervals instead of fixed (Step Average) (ablation) | ProfilingDiT foreground reuse · HunyuanVideo 720P | LPIPS: 0.1303 -> 0.1449 | Hurt |
| increase GPU count from 1 to 8 (ablation) | ProfilingDiT · HunyuanVideo 720P 129 frames | speedup: 1.87x -> 1.61x | Hurt |
| replace TeaCache with ProfilingDiT | HunyuanVideo · HunyuanVideo 720P (vs TeaCache slow) | PSNR: 21.91 -> 26.44 | Improved |
Closest papers by meaning
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.92)
- Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model (CVPR 2025 · similarity 0.91)
- Cache Me if You Can: Accelerating Diffusion Models through Block Caching (CVPR 2024 · similarity 0.90)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.89)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.88)
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.
