---
license: openrail
---

# Micro-World 2.0

Action-controlled causal world-model checkpoints. Micro-World 2.0 generates 352×640 video from a first frame, a text prompt, and keyboard/mouse actions. The distilled student denoises each chunk in 4 steps.

The Wan2.2-TI2V-5B base weights are not in this repository; download them from [Wan-AI/Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B).

## Checkpoints

| Path | What it is |
| --- | --- |
| `causal-forcing-init/model.pt` | Causal-forcing initialization. Loaded onto the Wan2.2 transformer before the LoRAs. |
| `causal-lora/lora_diffusion_pytorch_model.safetensors` | Causal LoRA. Adapts the backbone to open-domain 352×640 video. |
| `action/` | Action module. Inference reads `action_preprocess` and `action_adaLN_modulation` from this checkpoint. |
| `dmd-lora/lora_diffusion_pytorch_model.safetensors` | DMD student LoRA used for 4-step generation. |
| `light_vae/` | Optional TurboVAED decoder from [amd/HummingbirdXT](https://huggingface.co/amd/HummingbirdXT/tree/main/vae/wan22_v1_tiling_16_12). Faster decode with few loss of fidelity. Omit it to use the Wan VAE. |

## Load order

1. Wan2.2-TI2V-5B transformer, VAE, and text encoder
2. `causal-forcing-init/model.pt`
3. `causal-lora`
4. `action`
5. `dmd-lora`

Training and inference code: [Micro-World](https://github.com/AMD-AGI/Micro-World).