MiniMax H3 PTD LoRAs cut reference-to-video generation to 8 or 4 steps
Taobao Content AI released free acceleration LoRAs that make MiniMax H3's Ref2VA mode, which turns reference images into video with sound, run in 8 or 4 steps. A third, 8-step version trained with a method called RefAsymOPD claims sharper, more believable backgrounds at the same speed.
What it does
PTD (Parametric Trajectory Distillation, from EPFL, NVIDIA and Stanford researchers) trains one small adapter per sampling step, so the 8-step file runs in 8 model passes and the 4-step file in 4, and the step count is read from the file itself. The 8-step files are 11.1 GB and the 4-step file is 5.58 GB, all BF16, and the release ships a ComfyUI node pack and a Diffusers script.
Results and limits
The model cards show the authors' own side-by-side clips: the plain PTD files against the base Ref2VA model, the 4-step Turbo LoRA and the earlier 8-step PDD release, and the RefAsymOPD file against the 40-step base on five portrait clips (768×1344, 124 frames, 24 fps, with audio). No benchmark table is published.
Worth knowing
The official ComfyUI route needs the original, unpruned MiniMax H3 reference model converted with the included script, about 140 GB of weights in total, and ComfyUI started with --disable-dynamic-vram. The community ComfyUI-PTD-Compact nodes load the adapters as smaller INT8 files and on the pruned base. The LoRAs use the MiniMax H3 community licence, and a 4-step RefAsymOPD file is marked coming soon.