Peking University, Tsinghua and Alibaba Open-Source SparkDiffusion for Up to 265x Faster Wan Video Generation

Researchers from Peking University, Tsinghua University and Alibaba have open-sourced SparkDiffusion, an acceleration framework for Alibaba's Wan 2.1 and Wan 2.2 diffusion-transformer video models. Generating an 81-frame 720P video with Wan2.1-T2V-14B dropped from 4,769 to 18 seconds on a single RTX 5090 (265x), and from 1,757 to 8 seconds on an H100. Code is on GitHub under Apache 2.0, with checkpoints on Hugging Face.

Pandaily

Open source

81-frame 720P video on Wan2.1-T2V-14B: 4,769s to 18s on RTX 5090
Same job on H100 fell from 1,757s to 8s, a 220x speedup
Wan2.2-T2V-A14B at 720P: 4,545s to 25.1s, a 181x gain
Code on GitHub under Apache 2.0; 480P and 720P checkpoints released Sept 25

#Alibaba#Wan#SparkDiffusion#HuggingFace

Jensen Huang and Lisa Su join Tsinghua University advisory board

Stable AI and Tsinghua Release LimiX-2 Structured-Data Foundation Model

Infinigence AI Open-Sources APXInf for Embodied Edge Inference on Jetson Thor

Business

Peking University, Tsinghua and Alibaba Open-Source SparkDiffusion for Up to 265x Faster Wan Video Generation

Jensen Huang and Lisa Su join Tsinghua University advisory board

Stable AI and Tsinghua Release LimiX-2 Structured-Data Foundation Model

Infinigence AI Open-Sources APXInf for Embodied Edge Inference on Jetson Thor

Ericsson adapts wireless WAN control to sovereign shift