NVIDIA’s Sol-Attn speeds up video generation inference by 2.1x
Haopeng Li · hf · 2026-07-28
NVIDIA proposes Sol-Attn for faster video generation inference
The paper targets the attention bottleneck in diffusion transformers used for high-fidelity video generation. It introduces Sol-Attn (“Sparsifying online attention”), a training-free method that combines dynamic routing, sparse computation, and approximation correction in a single online-softmax pass.
What it changes
- Uses on-the-fly block thresholding instead of fixed top-k or cumulative-mass routing.
- Avoids materializing a proxy map by reusing proxy scores during routing.
- Approximates the contribution of dropped blocks instead of discarding them outright.
Reported results
- Improves the quality-efficiency trade-off for sparse attention.
- Delivers 2.1× end-to-end speedup for video generation.
- Delivers 2.3× end-to-end speedup for video editing.
- Evaluated across both image and video generation tasks while preserving visual quality.
Related event: NVIDIA Introduces Sol-Attn to Accelerate Video Generation(2 posts)→
More from Multimodal
- Flying car short made with Grok Imagine, Minimax H3 and Topaz upscale — creatoroff · 2026-10-16
- Fish Audio's open-weight Drama 3 voice model claims top emotive quality across 83 languages — mhdfaran · 2026-10-16
- Claude Opus 5.5 makes a pure-code Halloween animation starring bots trick-or-treating for tokens — sattvi · 2026-10-16
- 1,290-line commit lands MOSS Image support in ostris/ai-toolkit — ostrisai · 2026-10-16
- Ostris AI Toolkit adds training support for OpenMOSS's MOSS-Image-SFT model — ostrisai · 2026-10-16
- Nokia vs iPhone, car edition: AI image shows Tesla Roadster hovering over a Porsche — vaibhavbetter · 2026-10-16