Research map / Video / Video generation
TimeRipples: Accelerating vDiTs by Understanding the Spatio-Temporal Correlations in Latent Space
Wenxuan Mao, Yulin Sun, Aiyue Chen, Jing Lin, Yiwu Yao, Yiming Gan, Jieru Zhao, Jingwen Leng, Minyi Guo, Yu Feng
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 |
|---|---|---|---|
| use TimeRipple 75% reuse instead of SVG 70% masking | HunyuanVideo · HunyuanVideo | PSNR (dB): 25.78 -> 35.06 | Improved |
| use TimeRipple 85% instead of the original model | HunyuanVideo · HunyuanVideo | VBench (%): 80.28 -> 80.44 | No clear effect |
| raise reuse from 75% to 85% self-attention savings (ablation) | TimeRipple · HunyuanVideo | PSNR (dB): 35.06 -> 31.28 | Mixed |
| combine TimeRipple 75% with SVG 70% sparse masking (ablation) | TimeRipple 75% · HunyuanVideo | PSNR (dB): 35.06 -> 26.17 | Hurt |
| on Wan2.1, TimeRipple vs SVG on VBench | Wan2.1-T2V-14B · Wan2.1-T2V-14B | VBench (%): 81.47 -> 81.03 | Hurt |
| use adaptive per-step thresholds instead of a fixed threshold (ablation) | TimeRipple 75% · HunyuanVideo | PSNR (dB): 34.56 -> 35.06 | Improved |
| add spatial reuse in addition to temporal reuse (ablation) | TimeRipple 75% · HunyuanVideo | PSNR (dB): 32.44 -> 35.06 | Improved |
| set separate per-channel thresholds instead of a uniform threshold (ablation) | TimeRipple 85% · HunyuanVideo | VBench (%): 80.44 -> 79.37 | Hurt |
| use reuse over channels instead of masking tokens with same selection (ablation) | TimeRipple · attention output MSE | MSE | Improved |
| increase reuse window size from 2 to 4 or more (ablation) | TimeRipple 85% · HunyuanVideo | image quality / speedup | Hurt |
Closest papers by meaning
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.90)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.89)
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.89)
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.89)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.88)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · 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.
