Research map / Video / Video generation
Accelerating Text-to-Video Generation with Calibrated Sparse Attention
Shai Yehezkel, Shahar Yadin, Noam Elata, Yaron Ostrovsky-Berman, Bahjat Kawar
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 CalibAtt calibrated sparse attention instead of dense FA3 | Wan 2.1 14B 480p · Wan 2.1 14B 480p | End-to-end latency: 363s -> 250s | Improved |
| use CalibAtt instead of dense FA3 | Wan 2.1 14B 720p · Wan 2.1 14B 720p | End-to-end latency: 1244s -> 785s | Improved |
| use CalibAtt instead of dense FA3 | Mochi 1 480p · Mochi 1 480p | VBench Total: 74.54 -> 74.57 | No clear effect |
| use CalibAtt instead of dense FA3 on distilled 4-step model | LightX2V 720p · LightX2V 4-step 720p | End-to-end latency: 48.3s -> 30.6s | Improved |
| use CalibAtt instead of SVG2 | Mochi 1 480p · Mochi 1 480p | End-to-end latency: 237s -> 161s | Improved |
| use CalibAtt instead of SVG2 | LightX2V 4-step 480p · LightX2V 4-step 480p | End-to-end latency: 24.0s -> 11.2s | Improved |
| reduce calibration prompts from 64 to 16 mask prompts with 1 similarity prompt (ablation) | CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720p | calibration GPU-hours: 89.6 -> 13.7 | Improved |
| reduce mask calibration prompts to 8 (ablation) | CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720p | VBench Total: 79.69 -> 78.53 | Hurt |
| use 1 prompt instead of 64 for spatial similarity computation (ablation) | CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720p | calibration GPU-hours: 89.6 -> 53.0 | Improved |
| merge skip-list intervals to 90% plus timestep mask sharing tau 0.97 (ablation) | CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720p | mask memory (GB): 21.5 -> 3.6 | Improved |
| use 1D skip lists instead of 2D (ablation) | CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720p | mask memory (GB): 21.5 -> 4.0 | Improved |
| vary kernel block size from 128x64 to 128x192 (ablation) | CalibAtt Wan 2.1 14B 480p · Wan 2.1 14B 480p | speedup: 1.37x -> 1.42x | No clear effect |
Closest papers by meaning
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.89)
- LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 2025 · similarity 0.88)
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.88)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.88)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.88)
- Faster Video Diffusion with Trainable Sparse Attention (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.
