AI Paper Atlas

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

ECCV 2026 · Video › Video generation · approach: Compositional text-to-video diffusion · 3 citations (OpenAlex)

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 changedWhereMeasuredEffect
use CalibAtt calibrated sparse attention instead of dense FA3Wan 2.1 14B 480p · Wan 2.1 14B 480pEnd-to-end latency: 363s -> 250sImproved
use CalibAtt instead of dense FA3Wan 2.1 14B 720p · Wan 2.1 14B 720pEnd-to-end latency: 1244s -> 785sImproved
use CalibAtt instead of dense FA3Mochi 1 480p · Mochi 1 480pVBench Total: 74.54 -> 74.57No clear effect
use CalibAtt instead of dense FA3 on distilled 4-step modelLightX2V 720p · LightX2V 4-step 720pEnd-to-end latency: 48.3s -> 30.6sImproved
use CalibAtt instead of SVG2Mochi 1 480p · Mochi 1 480pEnd-to-end latency: 237s -> 161sImproved
use CalibAtt instead of SVG2LightX2V 4-step 480p · LightX2V 4-step 480pEnd-to-end latency: 24.0s -> 11.2sImproved
reduce calibration prompts from 64 to 16 mask prompts with 1 similarity prompt (ablation)CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720pcalibration GPU-hours: 89.6 -> 13.7Improved
reduce mask calibration prompts to 8 (ablation)CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720pVBench Total: 79.69 -> 78.53Hurt
use 1 prompt instead of 64 for spatial similarity computation (ablation)CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720pcalibration GPU-hours: 89.6 -> 53.0Improved
merge skip-list intervals to 90% plus timestep mask sharing tau 0.97 (ablation)CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720pmask memory (GB): 21.5 -> 3.6Improved
use 1D skip lists instead of 2D (ablation)CalibAtt Wan 2.1 14B 720p · Wan 2.1 14B 720pmask memory (GB): 21.5 -> 4.0Improved
vary kernel block size from 128x64 to 128x192 (ablation)CalibAtt Wan 2.1 14B 480p · Wan 2.1 14B 480pspeedup: 1.37x -> 1.42xNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy