FastH3 VSA DataFree LoRA

Updated

FastH3 (released by the FastVideo team at Hao AI Lab) is an accelerated, distilled version of the MiniMax H3 video-and-audio generation model. It achieves significant speedups—up to 14x on certain hardware—by reducing the required transformer inference calls from 49 to 4 and utilizing Video Sparse Attention (VSA) to prune approximately 90% of attention calculations.

Regarding "DataFree LoRA" in the context of FastH3, the term refers to the following:

What "Data-Free" Means for FastH3

In the FastH3 project, "Data-Free" denotes that the model was trained using "prompt-only" distillation (often leveraging the DMD2 framework) rather than requiring large sets of target videos. The recommended FastH3 checkpoint is explicitly titled "FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree".

The Role of LoRA in FastH3

  • Pre-extracted Adapter: The FastVideo team provides the FastH3 model in two main forms: full weights and a pre-extracted LoRA. If you use the LoRA version, you are essentially loading the distilled "FastH3" adaptation on top of the base MiniMax H3 model.
  • VSA Requirement: It is critical to note that the FastH3 LoRA is not a generic, drop-in PEFT (Parameter-Efficient Fine-Tuning) LoRA. It is a specific distilled adapter that requires the matching VSA-H3 backend and kernel to function. If you attempt to run this LoRA through a standard dense attention path (without the FastVideo VSA kernel), it will not produce correct outputs.
  • Usage: When using the FastVideo launcher or compatible inference pipelines (like those in ComfyUI), you select the VSA-DataFree checkpoint or its matching LoRA. The launcher then handles the necessary backend configurations to ensure the sparse attention kernels are active.

Summary of Specifications

  • Checkpoint: FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
  • Architecture: 4-step DMD2 distillation + 90% sparse VSA (tile size 64).
  • Compatibility: Requires the FASTVIDEO_VSA kernel and specific attention backends (e.g., Triton). It is currently designed for Text-to-Video-and-Audio (T2VA).