HEART framework boosts video diffusion model efficiency

Researchers have developed HEART, a novel training-free framework designed to enhance the efficiency of sparse attention mechanisms in video diffusion models. This method addresses head heterogeneity by adapting mask refreshing and threshold calibration based on observed head behavior across denoising steps. HEART leverages Temporal Mask Reuse (TMR) to intelligently update sparse masks and Error-guided Budgeted Calibration (EBC) to optimize thresholds, leading to improved quality-efficiency trade-offs without requiring model retraining or modifications. AI

IMPACT Improves efficiency in video diffusion models by optimizing sparse attention mechanisms.

RANK_REASON Research paper detailing a new method for improving video diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Error-guided Budgeted Calibration
• HEART
• HunyuanVideo-13B
• SVG2
• Temporal Mask Reuse
• Wan2.1-1.3B
• Wan2.1-14B
• XAttention
• Xuzhe Zheng

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →