PickFastVideo-FastH3-4-step-Preview-v1-VSA-DataFree: Simultaneous Video and Audio Generation with 4-Step Inference

Generate synchronized video and audio outputs from a single text prompt. A 35B parameter model based on MiniMax H3 is compressed using the DMD2 distillation technique, completing the generation process in just 4 Transformer forward passes.

Computational efficiency is improved by applying the VSA-H3 attention backend and 90% sparsity. While complex motion, fine details, and some audio quality may be lower compared to the original model, it is useful for rapid prototyping or in resource-constrained environments.

Default configurations tested on a 4x B200 GPU setup are provided. CUDA 13 and the Blackwell architecture are prioritized, while other multi-GPU environments can run via Triton kernels and specific flags.