MiniMax-H3 Acc LoRAs: few-step FL2VA and Ref2VA with Parallel Decoding Distillation

alibaba-pai applies Parallel Decoding Distillation (PDD; Shaul et al., arXiv 2607.26004, 2026) to MiniMax-H3 to cut generation to a few steps. Two LoRAs, both rank 64 with network_alpha 64 in BF16, match MiniMax-H3's two checkpoints: MiniMax-H3-FL2VA-Acc-8Step.safetensors for the FL2VA checkpoint (first and last keyframes) and MiniMax-H3-Ref2VA-Acc-8Step.safetensors for the Ref2VA checkpoint (reference conditioning). Use: in the VideoX-Fun repository, set model_path to the MiniMax-H3 model and pdd_lora_path to the matching LoRA in examples/MiniMax-H3-Acc-LoRAs/predict_t2v.py for FL2VA or predict_ref2v.py for Ref2VA, then run the script. Each script calls apply_pdd_lora, which loads the LoRA and derives the number of inference steps from its configuration. The scripts use diffusers' MiniMax-H3 ModularPipeline and need diffusers >= 0.40.0. Demos compare the base model, lightx2v's Minimax-h3-Turbo 4-step checkpoints and these LoRAs on test cases from the Minimax-H3-Turbo examples, generated at LoRA weight 1.0 with 4 and 8 NFE (function evaluations), FL2VA at 768p. The videos were not inspected here. Licence: MiniMax H3 Community License Agreement. Based on the alibaba-pai/MiniMax-H3-Acc-LoRAs card pinned to revision 335001fb9e5455d68a0caa18ec2e319072150328 and observed on 2026-10-27. Nothing was run here; the card names the LoRAs 8-step but its demos also use 4 NFE, and the step count chosen by apply_pdd_lora was not checked. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/raw/335001fb9e5455d68a0caa18ec2e319072150328/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 314:510, 672:1360, 1014:1360, 6397:6466, 6483:7014, 7090:7120; exact excerpt SHA256 4049e6626137c0126f06aab560788b009a5b0ffcb233db3c27de6f0d7af8e71b,7f825663c9dbd896d45738790801465e92612d239b0034450c9d1d3a4df0d7e7,b4c3ab8d08e97d5b14c14a6290cc16cbaba0a56d4eb3ec1a7023b0bfdc87aebf"
    }
  ]
}