Modelpublic

alibaba-pai/MiniMax-H3-Acc-LoRAs

sourceHugging Faceotherupdated 27d agoView on Hugging Face

242likes98kdownloads

Model Card

MiniMax-H3-Acc-LoRAs

Introduction

We apply Parallel Decoding Distillation (PDD) <sup>1</sup> 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)

<table border="0" style="width: 100%; text-align: center; margin-top: 20px;"> <thead> <tr> <th style="text-align: center;" width="33%">MiniMax-H3-FL2VA</th> <th style="text-align: center;" width="33%"><a href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimaxh3fl2vturbo4stepv1.1768pbf16.safetensors">Minimax-h3-Turbo <br> (fl2vturbo4stepv1.1768p)</a></th> <th style="text-align: center;" width="33%">MiniMax-H3-FL2VA-Acc-8Step </th> </tr> </thead> <tr> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vabaseline1.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vaturbo4stepv1.1768p1.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vaacc8stepv11.mp4" width="100%" controls loop></video> </td> </tr> <tr> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vabaseline2.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vaturbo4stepv1.1768p2.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vaacc8stepv12.mp4" width="100%" controls loop></video> </td> </tr> <tr> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vabaseline3.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vaturbo4stepv1.1768p3.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3fl2vaacc8stepv13.mp4" width="100%" controls loop></video> </td> </tr> </table>

Ref2VA

<table border="0" style="width: 100%; text-align: center; margin-top: 20px;"> <thead> <tr> <th style="text-align: center;" width="33%">MiniMax-H3-Ref2VA</th> <th style="text-align: center;" width="33%"><a href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimaxh3ref2vturbo4stepv0.1bf16.safetensors">Minimax-h3-Turbo <br> (ref2vturbo4stepv0.1)</a></th> <th style="text-align: center;" width="33%">MiniMax-H3-Ref2VA-Acc-8Step</th> </tr> </thead> <tr> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vabaseline1.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vaturbo4stepv0.1768p1.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vaacc8stepv11.mp4" width="100%" controls loop></video> </td> </tr> <tr> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vabaseline2.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vaturbo4stepv0.1768p2.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vaacc8stepv12.mp4" width="100%" controls loop></video> </td> </tr> <tr> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vabaseline3.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vaturbo4stepv0.1768p3.mp4" width="100%" controls loop></video> </td> <td> <video src="https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/results/minimaxh3ref2vaacc8stepv13.mp4" width="100%" controls loop></video> </td> </tr> </table>

[!NOTE] The above test cases are from <a href="https://github.com/ModelTC/Minimax-H3-Turbo/tree/main/examples">Minimax-H3-Turbo</a>. 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.

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

Reference

<ol> <li id="ref1">Neta Shaul, et al. "Parallel Decoding Distillation for Fast Image and Video Generation.". arXiv preprint arXiv:2607.26004 (2026).</li> </ol>