MiniMax-H3-Acc-LoRAs

Introduction

We apply Parallel Decoding Distillation (PDD) ¹ to MiniMax-H3, enabling efficient video generation in only a few inference steps.

For more details, please refer to our GitHub repo.

Name	Base Model	Hugging Face	Description
MiniMax-H3-FL2VA-Acc-8Step.safetensors	MiniMax-H3 (FL2VA)	🤗Link	Official 8 Step Acc LoRA (rank=64 and network_alpha=64 in BF16) for MiniMax-H3 (FL2VA).
MiniMax-H3-Ref2VA-Acc-8Step.safetensors	MiniMax-H3 (Ref2VA)	🤗Link	Official 8 Step Acc LoRA (rank=64 and network_alpha=64 in BF16) for MiniMax-H3 (Ref2VA).

Demo

FL2VA (768p)

MiniMax-H3-FL2VA	Minimax-h3-Turbo (fl2v_turbo_4step_v1.1_768p)	MiniMax-H3-FL2VA-Acc-8Step

Ref2VA

MiniMax-H3-Ref2VA	Minimax-h3-Turbo (ref2v_turbo_4step_v0.1)	MiniMax-H3-Ref2VA-Acc-8Step

The above test cases are from Minimax-H3-Turbo. Videos are generated with a LoRA weight of 1.0 at both 4 and 8 NFE.

Quick Start

Set model_path and pdd_lora_path to the MiniMax-H3 model and the matching acceleration LoRA checkpoint in predict_t2v.py for FL2VA or predict_ref2v.py for Ref2VA, then run the corresponding script. Each example uses apply_pdd_lora to load the checkpoint and derive the required number of inference steps from its configuration.

These scripts use Diffusers' MiniMax-H3 ModularPipeline and require diffusers >= 0.40.0.

Reference

1. Neta Shaul, et al. "Parallel Decoding Distillation for Fast Image and Video Generation.". arXiv preprint arXiv:2607.26004 (2026).