Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation
The paper claims a 14.5x denoising speedup for 14.3-second 768p video generation on eight B200 GPUs.
Video DeltaNet combines local Softmax attention with bidirectional linear memory to handle long-range video context without losing fine-grained interactions. Its Video Delta Attention branch updates memory once per frame using the frame’s spatial tokens, with gates and separate projections balancing the hybrid paths. The authors instantiate it on MiniMax H3, keeping Softmax for text and audio interactions while applying the hybrid path to video-to-video attention. With eight-step distillation and an optimized SGLang serving stack, VDN-H3 completes DiT denoising in 6.70 seconds. ArXiv · AI/CL/LG's note
Video DeltaNet combines local Softmax attention with bidirectional linear memory to handle long-range video context without losing fine-grained interactions. Its Video Delta Attention branch updates memory once per frame using the frame’s spatial tokens, with gates and separate projections balancing the hybrid paths. The authors instantiate it on MiniMax H3, keeping Softmax for text and audio interactions while applying the hybrid path to video-to-video attention. With eight-step distillation and an optimized SGLang serving stack, VDN-H3 completes DiT denoising in 6.70 seconds. ArXiv · AI/CL/LG's note
score 5