Prism model enables 2K video-audio generation with dynamic sparse attention

A new model named Prism has been released, designed for native joint video-audio generation at resolutions up to 2K. Developed by researchers from Fudan University and Tencent Hunyuan, Prism utilizes a dynamic sparse attention framework to efficiently process high-resolution data. This approach allows the model to adapt its attention structure based on local content, leading to a 2.5x training speedup compared to full attention while improving generation quality. AI

IMPACT This model's dynamic sparse attention could influence future approaches to efficient high-resolution generative AI training.

RANK_REASON The item describes the release of a new model and its associated technical report, detailing its architecture and performance. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Trending Models →

• Accelerate
• diffusers
• FrancisRing/Prism
• Fudan University
• Jiangfeng Xiong
• Jian Wei Zhang
• Kaihang Pan
• Prism
• PyTorch
• Qi Tian
• Shuyuan Tu
• Tencent Hunyuan
• transformers
• Weijie Kong
• Xintong Han
• Yinming Huang
• Yue Wu
• Yu-Gang Jiang
• Zuxuan Wu

• model release
• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →