Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation
VDN-H3 reports 14.3-second 768p video denoising in 6.70 seconds on eight NVIDIA B200 GPUs.
The paper proposes a hybrid attention design for video diffusion models, pairing local Softmax attention with bidirectional linear memory for longer video context. Its Video Delta Attention updates memory once per frame using the frame’s spatial tokens. The authors apply it to MiniMax H3 for video-to-video interactions, while keeping Softmax attention for text and audio interactions. They report a 14.5x speedup over a 50-step dense H3 baseline on the same GPU count. HF Daily Papers' note
The paper proposes a hybrid attention design for video diffusion models, pairing local Softmax attention with bidirectional linear memory for longer video context. Its Video Delta Attention updates memory once per frame using the frame’s spatial tokens. The authors apply it to MiniMax H3 for video-to-video interactions, while keeping Softmax attention for text and audio interactions. They report a 14.5x speedup over a 50-step dense H3 baseline on the same GPU count. HF Daily Papers' note
score 5