Video diffusion model cuts denoising time 14.5x with per-frame memory

single source· 1 articles · confidence: medium · first seen 2026-09-16 20:00 UTC

What this means for you

Nothing to deploy: no weights, no API, no release date. The headline speedup mixes a step-count reduction with the attention change, so the two are not separable from the numbers given. If you serve video diffusion, the transferable idea is keeping full attention where quality is fragile and a per-frame memory where it is not.

A 16 September arXiv paper describes Video DeltaNet, a way to cut the cost of video diffusion models, which generate clips by repeatedly refining noise. Denoising re-examines every part of every frame at each step. VDN keeps full attention — comparing every token with every other — for text and audio, and replaces it for video with a fixed-size memory updated once per frame. On MiniMax H3, with eight-step distillation and an SGLang serving stack, it denoises a 14.3-second 768p clip in 6.70 seconds on eight B200 GPUs, 14.5x faster than a 50-step dense baseline. That figure bundles fewer denoising steps with the attention change.

Key facts

  • ·Video DeltaNet combines local Softmax attention with bidirectional linear memory, with the linear branch updating memory once per frame. source
  • ·The method is instantiated on MiniMax H3, retaining Softmax attention for interactions involving text or audio. source
  • ·VDN-H3 denoises a 14.3-second, 768p video in 6.70 seconds on eight NVIDIA B200 GPUs. source
  • ·The paper reports a 14.5x speedup over a 50-step dense H3 baseline on the same GPU count. source
  • ·The reported result uses eight-step distillation and an optimised SGLang serving stack. source
  • ·The preprint was posted on 16 September 2026 as arXiv 2609.20744; no separate evaluation date is given for the speedup figure. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire