H3-World: Turning Language Understanding into World Control
H3-World is the first interactive world model built on MiniMax-H3. It generates action-controlled video from an initial frame by converting keyboard states into per-latent language instructions and binding each instruction to its corresponding future video latent through directed attention routing. Using 8,000 gameplay clips from ABot-World-Explorer-500h, H3-World learns 65.6M LoRA parameters, only 0.199% of the 33B backbone.
https://github.com/user-attachments/assets/1c862995-8809-447e-bade-2c47bfdb2738
H3-World: Turning Language Understanding into World Control
Danze Chen1,2, Zeqing Wang1,2, Ziyue Lin3, Xingyi Yang3, Yeying Jin1,2
1Tencent 2National University of Singapore 3The Hong Kong Polytechnic University
⚙️ Setup
Tested with Python 3.10 and CUDA 12.8.
conda create -n minimax_h3 python=3.10 -y
conda activate minimax_h3
pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt
# Use the exact DiffSynth revision and the H3-World attention patch.
git clone https://github.com/modelscope/DiffSynth-Studio.git DiffSynth-Studio-h3-v2
git -C DiffSynth-Studio-h3-v2 checkout "$(cat code/diffsynth_base_commit.txt)"
git -C DiffSynth-Studio-h3-v2 apply ../code/diffsynth_h3_action.patch
# Keep Hugging Face, Torch, and Triton caches inside this repository.
source env.sh
Download the required weights into the following locations:
| Asset |
Required for |
Target location |
| MiniMax-H3 FL2VA base weights |
inference and training |
DiffSynth-Studio-h3-v2/models/MiniMax/MiniMax-H3/ |
| H3-World LoRA |
inference |
checkpoints/H3-World/step-10000.safetensors |
| ABot-World-Explorer-500h |
training only |
any local path, passed through ABOT_SRC_ROOT |
python3 -c "
from huggingface_hub import snapshot_download
snapshot_download(
'MiniMaxAI/MiniMax-H3',
allow_patterns=['model_index.json', 'FL2VA/*'],
local_dir='DiffSynth-Studio-h3-v2/models/MiniMax/MiniMax-H3'
)"
python3 -c "
from huggingface_hub import hf_hub_download
hf_hub_download(
'DANNY621/H3-World',
'step-10000.safetensors',
local_dir='checkpoints/H3-World'
)"
H3-World uses the original FL2VA checkpoint layout released by MiniMax, rather than the root-level Diffusers checkpoint. No weight conversion or key remapping is required.
The patch is required for the released checkpoint. Do not install DiffSynth-Studio-h3-v2 in editable mode; the included training and inference scripts verify that they load the patched checkout.
🎬 Inference
The repository includes a held-out ABot test frame at examples/first_frame.png. The following fixed configuration generates a 5.2-second forward-motion video:
python3 code/abot/infer.py \
--checkpoint checkpoints/H3-World/step-10000.safetensors \
--first-frame examples/first_frame.png \
--scene-prompt "A man in a yellow floral shirt stands in a dim, multi-level concrete parking garage." \
--action-preset forward \
--seed 2 \
--steps 50 \
--num-frames 124 \
--cfg-scale 1.0 \
--out outputs/example_forward.mp4
The included frame is sample d0b768c6 from the held-out test split. To use a custom image, replace examples/first_frame.png and describe its static scene and subject with --scene-prompt. Inputs are center-cropped to 832x480 when needed. The built-in presets are still, forward, back, strafe-left, strafe-right, tilt-up, tilt-down, pan-left, pan-right, pan-left-fast, and pan-right-fast. The full key-to-language mapping is defined in code/abot/action_script.py.
🏋️ Training
Prepare the 7,872-clip training split, cache its latents, inject the per-latent action instructions, then train LoRA:
# 1. Build clips and the fixed train/test split from ABot.
ABOT_SRC_ROOT=/path/to/ABot-World-Explorer-500h \
python3 code/abot/build_abot_clips.py --num-clips 8000 --workers 48
ABOT_SRC_ROOT=/path/to/ABot-World-Explorer-500h \
python3 code/abot/build_abot_clips.py --verify 8
python3 code/abot/split_abot_metadata.py \
--input data/abot_meta_8000.jsonl \
--train-output data/abot_meta_train_7872.jsonl \
--test-output data/abot_meta_test_128.jsonl \
--clips-dir data/clips
# 2. Cache the video, audio, and text latents, then add action text.
bash code/cache.sh
python3 code/abot/inject_abot_text.py \
--meta data/abot_meta_train_7872.jsonl \
--cache output/minimax_h3_abot/7872-cache \
--device cuda:0
# 3. Train on four GPUs by default.
bash code/train.sh
code/train.sh uses rank-32 LoRA on qkv_proj and out_proj for 20 epochs, saving checkpoints every 2,000 steps. Override the visible devices with CUDA_VISIBLE_DEVICES=4,5,6,7 bash code/train.sh.
🙏 Acknowledgements
📚 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},
}
H3-World 是一个基于 MiniMax-H3 构建的交互式世界模型:给它一帧初始画面,它就会生成你可以用键盘操作来控制的视频,就像用按键开车一样在游戏世界里驰骋。