MiniMax H3-World for ComfyUI

This repository provides the released H3-World action-control LoRA in an install-ready ComfyUI folder layout.

The checkpoint was created by the H3-World authors. It has not been converted, merged, quantized, or otherwise modified by T8star. The uploaded file is byte-for-byte identical to DANNY621/H3-World at revision 8883340a740842b5c40b45383e5b75df34931806.

Use it with the ComfyUI MiniMax H3 Audio T8 custom nodes. The node package supplies the H3-World action timeline, directed-attention integration, first-frame I2VA conditioning, and safe H.264/AAC output workflow required by this LoRA.

Included file

loras/minimax/H3-World/step-10000.safetensors

File	Size	SHA-256
`step-10000.safetensors`	131,227,832 bytes	`ddd9187b920b1e52c2d090f4e264fd83d8d433efc2a5b159e58883aeaf96e526`

It is the original rank-32 LoRA with 208 tensors (104 A/B pairs) and can be loaded directly by ComfyUI. No converted checkpoint is needed.

Install

Run this command from the directory that contains your ComfyUI folder:

hf download t8star/Minimax-H3-World-Comfy --include "loras/**" --local-dir ComfyUI/models

After downloading, the final path must be:

ComfyUI/models/loras/minimax/H3-World/step-10000.safetensors

Restart ComfyUI after installing the file.

Required ComfyUI workflow

Install or update comfyui-minimax-h3-audio-T8, then open:

examples/workflows/26-h3-world/2026-09-04_H3_World_I2VA_832x480_124f_50step_Advanced.json

The first release uses this fixed generation contract:

first-frame I2VA
832×480, 124 frames, 24 fps
50 Euler/native-flow steps, CFG 1.0
37 action latents
W/A/S/D for character movement
I/J/K/L, with optional F, for camera movement

The workflow also needs the existing full MiniMax H3 FL2VA base model, Qwen3-VL text encoder, video VAE, and audio VAE. Those files are not duplicated in this repository. Follow the model table in the node README for their locations.

Do not use this LoRA with an unmodified generic LoRA workflow. H3-World requires its per-action text encoding, mirrored time positions, and directed attention routing. The T8 workflow provides those pieces.

Provenance and license

Original model: DANNY621/H3-World
Official code: Danzer1xxxxChan/H3-World
Mirrored model revision: 8883340a740842b5c40b45383e5b75df34931806
ComfyUI integration: T8mars/comfyui-minimax-h3-audio-T8

The H3-World LoRA is released under Apache 2.0. MiniMax H3 base weights remain subject to the MiniMax H3 license and are not included here.

Citation

@misc{chen2026h3worldturninglanguageunderstanding,
  title={H3-World: Turning Language Understanding into World Control},
  author={Danze Chen and Zeqing Wang and Ziyue Lin and Xingyi Yang and Yeying Jin},
  year={2026},
  eprint={2609.01560},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.01560}
}

t8star/Minimax-H3-World-Comfy