---
library_name: pytorch
license: other
license_name: mixed-mit-and-apache-2.0
license_link: https://github.com/CompVis/WhatMoves/blob/main/THIRD_PARTY_NOTICES.md
pipeline_tag: image-to-video
tags:
- motion-representation
- motion-transfer
- wan2.2
---

# WhatMoves checkpoints

Released checkpoints for [WhatMoves](https://github.com/CompVis/WhatMoves), a
promptable model for localized motion representations and transfer.

**Paper:** [What Moves? Localized Motion Representations for Compositional Scene Control](https://huggingface.co/papers/2609.04383)  
**Project page:** [https://compvis.github.io/WhatMoves](https://compvis.github.io/WhatMoves)

| File | Contents |
|---|---|
| `what_moves/gated_static_step600000.pt` | Compact inference checkpoint for the WhatMoves motion encoder |
| `wan/gated_static_step600000/adapter_step30500.pt` | Motion projector and LoRA deltas for Wan2.2 I2V-A14B |
| `wan/gated_static_step600000/metadata.json` | Architecture, provenance, and SHA-256 metadata |

The source repository provides the model definitions and Torch Hub entrypoints;
these files are not intended to be loaded with a generic Diffusers pipeline.

## Licensing

This repository contains mixed-license model artifacts:

- Original WhatMoves parameters are released under the MIT License.
- The WhatMoves checkpoint incorporates parameters initialized from
  [DINOv2](https://github.com/facebookresearch/dinov2), licensed under Apache
  License 2.0.
- The motion-transfer adapter is designed to operate over
  [Wan2.2 I2V-A14B](https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B), whose code
  and base weights are licensed under Apache License 2.0. The base Wan weights
  are not redistributed here.

The MIT license does not supersede the licenses or notices of the upstream
components. See the source repository's `LICENSE`, `THIRD_PARTY_NOTICES.md`,
and bundled Apache License 2.0 text for the complete notices.