Video diffusion model adapted to edit footage live at 15 FPS

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

What this means for you

Nothing to act on yet: this is a paper plus a repository, not a service, with no pricing, no availability date and no independent evaluation. If you build video pipelines, the two claims worth testing yourself are the 15 frames per second and the single-H100 requirement, both the authors' own.

SVEET, a framework posted to arXiv, edits video as it streams in rather than requiring the whole clip first. It adapts a pretrained bidirectional video diffusion model (one that normally sees every frame at once) by training an auxiliary branch with temporally independent self-attention, so control features can be injected frame by frame. The authors report 15 frames per second on a single Nvidia H100 without any acceleration tricks, and say editing quality holds up; no evaluation date or harness is given, and the comparison figures are their own. Code is on GitHub.

Key facts

  • ·SVEET edits video in an auto-regressive, streaming fashion by adapting a pretrained bidirectional video diffusion model rather than training a new one. source
  • ·The authors report 15 frames per second on a single Nvidia H100 GPU with no auxiliary acceleration techniques. source
  • ·The method is built on two stated principles: backbone feature disentanglement and conditional frame independence. source
  • ·An auxiliary model branch encodes source video with temporally independent self-attention and injects intermediate features into corresponding backbone blocks. source
  • ·Code is released at github.com/YujiaHu1109/SVEET. source
  • ·The paper is arXiv 2609.24788, posted 20 September 2026. source

What the sources say

Sources

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

← the wire