Update README.md

README.md

@@ -4,6 +4,7 @@ tags:
 - video
 - image-to-video
 - world-model
+- interactive-world-model
 - lora
 - diffusion
 - action-conditioning
@@ -11,55 +12,59 @@ tags:
 base_model: MiniMax/MiniMax-H3
 ---
 
-# H3-World
+# 🎮 H3-World: Turning Language Understanding into World Control
 
-Action-controllable world model built by fine-tuning [MiniMax-H3](https://huggingface.co/MiniMax/MiniMax-H3)
-on [ABot-World-Explorer-500h](https://huggingface.co/datasets/acvlab/ABot-World-Explorer-500h) gameplay
-footage. The same first frame plus a different key sequence produces a different, correct motion:
-`W A S D` for character movement, `I J K L` for camera tilt/pan, plus a synthesized `F` (fast) bit.
+H3-World is the **first interactive world model** built on
+[MiniMax-H3](https://huggingface.co/MiniMax/MiniMax-H3). Given an initial frame
+and keyboard controls, it generates action-controlled video with coordinated
+character and camera motion.
 
-![teaser](teaser.png)
+H3-World converts keyboard states into one language instruction per future video
+latent and uses directed attention routing to bind each instruction to its
+corresponding latent interval. Using 8,000 gameplay clips from
+[ABot-World-Explorer-500h](https://huggingface.co/datasets/acvlab/ABot-World-Explorer-500h),
+it learns 65.6M LoRA parameters, only 0.199% of the 33B MiniMax-H3 backbone.
 
-## What's in this repo
 
-| File | Description |
-|---|---|
-| `step-10000.safetensors` | LoRA checkpoint, step 10,000 of the released training run |
-
-## Method
-
-A pretrained video model already carries a language-to-motion prior from its training data, so
-rather than learning action control from scratch, we ground keyboard input in that existing prior.
-Key presses are converted into one short English sentence per latent video frame (e.g. *"the man
-walks forward, camera pans left sharply"*) and injected through MiniMax-H3's native text channel —
-the same channel it was already pretrained on. A directed attention mask binds each sentence to its
-own video frame so the model knows which instruction applies when. The only thing trained is a
-rank-32 LoRA on `qkv_proj`/`out_proj` (65.6M parameters, 0.198% of the 33.1B-parameter DiT).
+## Checkpoint
 
-Trained on 7,872 clips (832x480, 124 frames / 5.2s @ 24fps) from ABot-World-Explorer-500h.
+| File | Description |
+| --- | --- |
+| `step-10000.safetensors` | Released rank-32 H3-World LoRA checkpoint |
 
-## How to use
+Character controls use `W`, `A`, `S`, and `D`. Camera controls use `I`, `J`,
+`K`, and `L`, with `F` indicating fast camera motion.
 
-This checkpoint is a LoRA delta for MiniMax-H3 and requires the base model plus a directed-mask
-attention patch to run (the mask is what binds each sentence to its frame; without it the LoRA
-weights alone don't reproduce the reported behavior).
+## Usage
 
-```python
-from safetensors.torch import load_file
+This checkpoint is a LoRA delta for MiniMax-H3. It requires the MiniMax-H3 base
+weights and H3-World's directed-attention patch; loading the LoRA through an
+unmodified MiniMax-H3 pipeline will not reproduce the reported behavior.
 
-state = load_file("step-10000.safetensors")
-lora_state = {k: v for k, v in state.items() if ".lora_A." in k or ".lora_B." in k}
-# pipe.load_lora(pipe.dit, state_dict=lora_state, hotload=True)
-```
+Please use the official [H3-World code repository](https://github.com/Danzer1xxxxChan/H3-World)
+for setup, inference, and training instructions.
 
 ## License
 
-The LoRA weights in this repo are released under Apache 2.0. Using them requires the base
-[MiniMax-H3](https://huggingface.co/MiniMax/MiniMax-H3) model, which is governed by its own
-license terms.
+The H3-World LoRA checkpoint is released under Apache 2.0. MiniMax-H3 is
+governed by its own license terms.
+
+## Citation
+
+```bibtex
+@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},
+}
+```
 
 ## Acknowledgements
 
-- [MiniMax-H3](https://huggingface.co/MiniMax/MiniMax-H3) — the base model
-- [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio) — the training framework
-- [ABot-World-Explorer-500h](https://huggingface.co/datasets/acvlab/ABot-World-Explorer-500h) — the gameplay data
+- [MiniMax-H3](https://huggingface.co/MiniMax/MiniMax-H3)
+- [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio)
+- [ABot-World-Explorer-500h](https://huggingface.co/datasets/acvlab/ABot-World-Explorer-500h)