LightEWM

LightEWM: Light Embodied World Model is a lightweight, hackable training and inference codebase for embodied world models. It covers both bidirectional video diffusion workflows and causal world-modeling workflows, with examples for LIBERO and custom robot datasets.

The project is designed to be easy to modify: datasets, model backends, training configs, and inference entrypoints are kept small and explicit. For custom data, LightEWM provides a simple CSV-based interface so users can start training from lightweight metadata without building a complex dataset pipeline.

We also provide multiple pretrained checkpoints for fast post-training and downstream adaptation, including robot-pretrained initialization weights and LIBERO-finetuned models. The current main model family is Wan2.2-TI2V-5B, with support for LIBERO preprocessing, latent-cache generation, full training, and inference.

🧭 Roadmap

• Wan2.2-TI2V-5B training and inference support
• LIBERO preprocessing, training, and inference support
• Causal video model adaptation for Wan 5B
• WAM actor implementation for Wan 5B
• CALVIN preprocessing, training, and inference support
• Reinforcement learning on top of the world model

🛠️ Installation

1) Environment setup

git clone https://github.com/XuWuLingYu/LightEWM.git
cd LightEWM
conda create -n lightewm python=3.10
conda activate lightewm
pip install -e .
pip install flash-attn --no-build-isolation

2) Download checkpoint

hf download Wan-AI/Wan2.2-TI2V-5B \
  --local-dir ./checkpoints/Wan2.2-TI2V-5B

# Robot-pretrained DiT init checkpoint used by LIBERO/CustomDataset 5B training
hf download XuWuLingYu/Wan2.2-5B-Robot \
  --local-dir ./checkpoints/Wan2.2-5B-Robot

This downloads the full Hugging Face repository to ./checkpoints/Wan2.2-TI2V-5B, which matches the default path used by the TI2V-5B example configs under examples/LIBERO/. XuWuLingYu/Wan2.2-5B-Robot is pretrained on broad robot datasets and is the recommended initialization for downstream finetuning.

🚀 Quick Start on LIBERO

See LIBERO example. For causal world-model training and inference, see LIBERO-Causal example. For FVD/SSIM/LPIPS checkpoint comparison and smoke checks, see video quality evaluation.

📦 Run on Custom Dataset

See CustomDataset example.

🗂️ Model Zoo

• XuWuLingYu/Wan2.2-5B-Libero https://huggingface.co/XuWuLingYu/Wan2.2-5B-Libero/tree/main LIBERO-pretrained checkpoint trained on the full LIBERO dataset, suitable for direct LIBERO inference.
• XuWuLingYu/LIBERO-Causal-Wan2.2-5BTI2V https://huggingface.co/XuWuLingYu/LIBERO-Causal-Wan2.2-5BTI2V LIBERO-Causal checkpoint adapted from Wan2.2-TI2V-5B for causal robot video generation and direct LIBERO-Causal inference.
• XuWuLingYu/Wan2.2-5B-Robot https://huggingface.co/XuWuLingYu/Wan2.2-5B-Robot Pretrained on broad robot datasets; recommended as the initialization checkpoint for finetuning on LIBERO/CustomDataset.

📚 Reference

• DiffSynth-Studio: https://github.com/modelscope/DiffSynth-Studio
• Wan documentation: https://diffsynth-studio-doc.readthedocs.io/en/latest/Model_Details/Wan.html

📄 License

Apache-2.0

📖 Cite Us

If you find LightEWM useful for your research or applications, please cite:

@misc{lightewm2026,
  title = {LightEWM: Light Embodied World Model},
  author = {Community, LightEWM},
  year = {2026},
  url = {https://github.com/XuWuLingYu/LightEWM}
}