VideoDeltaNet

Video DeltaNet (VDN) is a hybrid attention architecture for video generation models. It replaces most of the quadratic softmax attention in MiniMax H3 with a mix of local sliding-window softmax attention and a long-range linear branch ("Video Delta Attention"), which turns the token-wise delta rule into a frame-wise one. The resulting checkpoint, VDN-H3, is distilled to 8 steps. The authors are from UC Berkeley, Impossible, Inc. and UT Austin.

The softmax branch attends between neighbouring chunks of latent frames and adds boundary anchors on the first and last frames. Distant context goes through bidirectional linear attention with forward and reverse states, and the two branches are combined with learned sigmoid gates. According to the project page, an 8×B200 setup with SGLang Diffusion produces a 14.4-second 768p video in about 9.0 seconds after warm-up, which is faster than playback. It reports near-lossless quality relative to the original MiniMax H3 and better quality and instruction following than MiniMax FastH3. One checkpoint supports text-to-video and first-frame, last-frame and first-and-last-frame conditioned generation, with audio.

The reference implementation is released under Apache-2.0. The weights are distributed under the MiniMax H3 Community License, which the ComfyUI port's README says excludes some territories and should be read before use. A community ComfyUI port (ComfyUI-VDN-H3) runs the model with portable PyTorch kernels on consumer GPUs. It is slower than the datacenter setup and needs considerable VRAM, and its author does not recommend it for limited hardware. The headline speedups require a multi-GPU B200 cluster.