Veda from ByteDance, HKU, and USTC: 5x Faster Video Diffusion

Teams from ByteDance, HKU, and USTC have open-sourced Veda, a sparse attention method for video diffusion transformers that accelerates video generation on Waver-T2V-12B from 19.4 to 3.8 minutes while preserving quality.

Compiled by Sergey KostenchukPublished 2026-09-27Updated 2026-09-27

2026-09-27 Research HuggingFace

🤖 Veda from ByteDance, HKU, and USTC: 5x Faster Video Diffusion

ByteDance, HKU, and USTC have open-sourced the code for Veda, a sparse attention method for video diffusion transformers (ICML 2026, MIT, fp8 predictor on Hugging Face). A lightweight predictor selects the necessary key tiles, and a block-sparse kernel computes only those: on Waver-T2V-12B, generation was accelerated from 19.4 to 3.8 minutes — 5.1x at 95% sparsity.

🌍 At 90-95% sparsity, quality is determined not by the percentage of discarded tokens, but by the accuracy of reproducing the dense attention structure: the mask is distilled through KL divergence, and tiles are selected for each head and layer.

👤 The code is reproducible on your own hardware: instructions cover RTX 4090, Hopper, and Blackwell, and scripts/generate.py compares dense and Veda versions with timings. On an RTX 4090, a 14.4-second clip for MiniMax-H3 runs 3.1 times faster than dense attention.

Source 1: https://veda-sparse.github.io/ Source 2: https://github.com/veda-sparse/Miowtion

Sources