PKU, Tsinghua, Alibaba open-source SparkDiffusion: 265x faster AI video on one RTX 5090
lmoroney · x · 2026-10-01
Researchers from Peking University, Tsinghua, and Alibaba open-sourced SparkDiffusion, cutting 720p Wan2.1 14B video generation from 4,769s to 18s on a single RTX 5090 — about 265x faster — using sparse attention, few-step distillation, and FP8. Code and weights are public. The author notes speed mainly helps iteration; taste, continuity, and story remain the real craft questions.
More from Multimodal
- Dev uses AI render projection to build Silent Hill-style Three.js kitchen with Astra — nptacek · 2026-10-01
- Ideogram's new Edit Model lands as users try AI outfit swaps — lishali88 · 2026-10-01
- Reddit user replicates Viggle-style motion transfer and character swap locally — Useful_Ad_52 · 2026-10-01
- Ideogram 4.5 lands on Magnific for all plans, with web, desktop and MCP support — aziz4ai · 2026-10-01
- Runway Research Q&A: Real-Time Models, Generative Interfaces and What's Next for Robotics — runwayml · 2026-10-01
- NVIDIA's LongLive-Plug distills video generation into plug-and-play LoRAs for 54 models — _akhaliq · 2026-10-01