Add model card

README.md

@@ -0,0 +1,70 @@
+---
+license: apache-2.0
+tags:
+- video
+- image-to-video
+- world-model
+- lora
+- diffusion
+- action-conditioning
+- game-world-model
+base_model: MiniMax/MiniMax-H3
+---
+
+# H3-World
+
+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.
+
+![teaser](teaser.png)
+
+## What's in this repo
+
+| File | Description |
+|---|---|
+| `step-10000.safetensors` | LoRA checkpoint, step 10,000 of the released training run |
+| `teaser.png` | Figure shown above |
+
+## Method
+
+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 the model was already pretrained on. A **directed attention mask** binds annotation
+*k* to video frame *k* only: frame *k* can read annotation *k*, but annotation *k* cannot read any
+other annotation, which closes a bypass a symmetric mask would leave open. Removing the annotation
+rows recovers the base model exactly.
+
+This checkpoint is a **rank-32 LoRA on `qkv_proj`/`out_proj`** across the 50 DiT blocks and 2
+token-refiner layers (104 modules, 208 tensors, 65.6M parameters — 0.198% of the 33.1B-parameter
+DiT). No action-specific module is added; the condition path itself carries zero extra weights.
+
+Trained on 7,872 clips (832x480, 124 frames / 5.2s @ 24fps) from ABot-World-Explorer-500h.
+
+## How to use
+
+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). Code, the framework patch, and a minimal
+inference script are not published alongside this checkpoint yet.
+
+```python
+from safetensors.torch import load_file
+
+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)
+```
+
+## 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.
+
+## 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