NVIDIA’s Sol-Attn delivers up to 5.08× end-to-end speedup for video generation
songhan_mit · x · 2026-07-28
Sol-Attn speeds up video generation with training-free sparse attention
NVIDIA Research introduced Sol-Attn, a training-free sparse attention method that accelerates pretrained visual generators while preserving quality.
- It combines dynamic routing, sparse computation, and approximate correction in a single online-softmax pass.
- The method uses on-the-fly block thresholding for controllable budgets and reuses proxy scores to approximate skipped blocks.
- Reported speedups vs. dense FlashAttention-3:
- Wan 2.1-14B: 2.02× end-to-end
- HunyuanVideo-13B: 2.12× end-to-end
- LTX 2.3: up to 2.4× end-to-end
- Integrated into Sol-Engine with kernel fusion and caching, the gains increase further:
- Wan 2.1-14B: 3.48× end-to-end
- HunyuanVideo-13B: 5.08× end-to-end
NVIDIA says the technique is already available in Sol-Engine, and the B200 kernel is still being optimized. The paper and code are public.
Related event: NVIDIA Introduces Sol-Attn to Accelerate Video Generation(2 posts)→
More from Multimodal
- MiniMax-generated horror series goes viral: the toy moves every time you blink — fa6637 · 2026-10-16
- AI-generated series "The Dining Car" Ep 4: the world ended, dinner service didn't — Acceptable-Day5621 · 2026-10-16
- Midjourney trick: layer a --profile moodboard over structured prompts for cinematic dark looks — LudovicCreator · 2026-10-16
- Multigen launches: a frontend for GPT-Image to build consistent image collections — Markasp · 2026-10-16
- One-prompt Skill generates an 18-second glitch-style product launch video — yihui_indie · 2026-10-16
- FLUX 3 vs Krea 2 vs Z-Image Turbo vs Qwen 2.1: Four-Way Image Model Comparison — Ok_Contribution8157 · 2026-10-16