Veda

Veda is a learned sparse-attention method for video diffusion that reduces attention computation by skipping low-importance tiles. It achieves 90% sparsity while maintaining perceptual quality, accelerating 720P video generation by 5.1× end-to-end (19.4 min to 3.8 min on 241 frames). The method uses a distilled predictor to identify which attention tiles matter most, then computes only the top 10% of them.

Veda ships as a ComfyUI plugin for MiniMax-H3 video models (T2VA, FL2VA, R2VA). It works as a plug-and-play attention override compatible with LoRAs, quantized checkpoints, and reference conditioning. Built with Claude Code and presented at ICML 2026.