EA-WM

Official final checkpoint for EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields.

• Paper: https://arxiv.org/abs/2605.06192
• Code and documentation: https://github.com/Shownx-c/EA-WM
• Base model: https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B

Model

Given an initial RGB frame and a text instruction, EA-WM jointly generates future RGB and KVAF videos through event-aware bidirectional fusion.

Checkpoint

File	Contents	Use
final.safetensors	RGB-stream LoRA, KVAF-stream LoRA, KVAF prediction head, and event-aware bidirectional fusion	Inference and evaluation

The checkpoint contains EA-WM adapters and auxiliary branch weights. Download the Wan2.2-TI2V-5B base model separately. Follow the repository README to install the code, prepare the base model, and run inference.

Download

hf download shown21/EA-WM final.safetensors --local-dir weights

Inference

From the EA-WM code directory, create a task list using configs/inference_tasks.example.json, then run:

export CHECKPOINT=weights/final.safetensors
export TASK_JSON=configs/inference_tasks.example.json
export OUTPUT_ROOT=outputs/inference/joint_generation

bash scripts/infer.sh

RGB videos are saved in videos/ and jointly generated KVAF videos in kvaf_videos/ under OUTPUT_ROOT.

Checkpoint metadata

The checkpoint uses the public EA-WM parameter names, LoRA rank 32, and fusion interval 5. File integrity can be checked against SHA256SUMS.

Citation

@article{yang2026eawm,
  title   = {EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields},
  author  = {Yang, Zhaoyang and Jin, Yurun and Qi, Lizhe and Chen, Kai and Huang, Cong},
  journal = {arXiv preprint arXiv:2605.06192},
  year    = {2026}
}

License

The EA-WM code and released checkpoint are provided under Apache-2.0. Use of the Wan2.2 base model and datasets remains subject to their respective terms.