Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation

Source facts

Video DeltaNet combines local Softmax attention with bidirectional linear memory for long-range video context. Its Video Delta Attention updates memory once per frame, with separate projections and gates for calibration. The preprint instantiates the design on MiniMax H3 and reports 14.3-second 768p video denoising in 6.70 seconds on eight NVIDIA B200 GPUs after eight-step distillation, compared with a reported 50-step dense baseline.

Analysis

The work targets the attention bottleneck in long video diffusion by assigning local and long-range interactions different computational paths. If independently validated, this is an infrastructure-level route to reducing generation latency without treating all video-token interactions as equally expensive.

Uncertainty/Doubt

• This is an arXiv preprint; the speedup, quality trade-offs, and hardware comparison were not independently reproduced here.