Taobao Slashes MiniMax-H3 Video Generation by 10x With Specialized Adapters
Taobao's Content AI team squeezes MiniMax-H3 reference-to-video-plus-audio generation into 4 or 8 denoising steps using per-hop LoRA adapters.
Takeaways
- Taobao Content AI released MiniMax-H3-PTD-LoRAs, cutting Ref2VA inference to 4 or 8 steps.
- Uses Parametric Trajectory Distillation: one independent LoRA adapter per denoising step.
- Built on top of Parallel Decoding Distillation's per-interval output heads.
- Renders 1344x768, 243 frames at 24 fps with 32 kHz stereo audio.
- Works in Diffusers 0.40+ and ships ComfyUI custom nodes; base must be unpruned Ref2VA transformer.
- A RefAsymOPD sibling release promises better sharpness and background realism at 6 or 8 steps.
Taobao’s content AI team has released two acceleration adapters for MiniMax-H3, an open-weight model that generates video with synchronized audio from a reference image. The adapters reduce its Ref2VA pipeline from 40 transformer evaluations to eight or four while leaving the base model’s weights unchanged.
The checkpoint repository provides both variants as BF16 safetensors. Each uses rank 128 LoRAs trained with Parametric Trajectory Distillation, or PTD.
A specialist adapter for every hop
PTD assigns an independent low-rank adapter to each inference hop instead of sharing one acceleration LoRA across the entire denoising trajectory. A LoRA adds a small set of trained weight updates to a frozen model, reducing the storage and training cost of adapting the full network.
The method builds on Parallel Decoding Distillation, or PDD, which supplies separate output heads for timestep intervals. During inference, each hop selects its own LoRA and fuses the corresponding PDD head. The eight-hop checkpoint therefore runs eight transformer evaluations, while the four-hop checkpoint runs four.
Step-specific weights give each adapter a narrower part of the trajectory to learn. Taobao’s published comparisons claim this design preserves motion, identity, lip synchronization, and audio structure better than a shared PDD LoRA at matched compute.