Researchers present Video DeltaNet (VDN), a hybrid attention architecture for video diffusion models that combines local softmax attention with a bidirectional linear-memory branch, Video Delta Attention, to address the computational bottleneck of processing long spatiotemporal token sequences during denoising. A staged teacher-alignment recipe progressively introduces the new linear pathway into pretrained video models while retaining softmax attention for text and audio interactions. Instantiated on MiniMax H3 with eight-step distillation and an optimized SGLang serving stack, VDN-H3 generates a 14.3-second 768p video in 6.70 seconds on eight NVIDIA B200 GPUs, a 14.5x speedup over the 50-step dense baseline.
