# PhysDreamer Reference Assets

Source: [YunjinZhang/PhysDreamer](https://huggingface.co/datasets/YunjinZhang/PhysDreamer)
Paper: [PhysDreamer: Physics-Based Interaction with 3D Objects via Video Generation](https://arxiv.org/abs/2406.04338)
Licence: MIT

## Contents

### `scenes/`

Four real-world scenes with RGB-D observations and optimized 3D Gaussians:

| Scene | Description |
|---|---|
| `alocasia` | Potted plant (flexible leaves) |
| `carnation` | Carnation flower (bendable stem) |
| `telephone` | Telephone handset (rigid + soft cord) |
| `hat` | Woven hat (flexible brim) |

Each scene directory contains:
- `point_cloud/iteration_*/point_cloud.ply` — 3DGS model (use as `--gaussians` input to `squishy-build`)
- `images/` — training RGB frames
- `cameras.json` — camera extrinsics and intrinsics

### `models/`

Optimized velocity fields and material fields for each scene:
- `velocity_field.pt` — PhysDreamer learned velocity field (per-particle trajectory prior)
- `material_field.pt` — optimized material parameters {E, nu, rho} per spatial region

## Usage with squishy

```python
# Use as gaussians input for squishy-build:
# squishy-build --gaussians reference/physdreamer/scenes/carnation/point_cloud/ \
#               --mesh carnation_mesh.obj --name carnation

# Use material field as initial parameter estimate for system ID:
from squishy.train.system_id import run_system_id_physics3d
# load material_field.pt to initialize E_init, nu_init
```

## Citation

```bibtex
@article{zhang2024physdreamer,
    title={PhysDreamer: Physics-Based Interaction with 3D Objects via Video Generation},
    author={Tianyuan Zhang and Hong-Xing Yu and Rundi Wu and
            Brandon Y. Feng and Changxi Zheng and Noah Snavely and Jiajun Wu and William T. Freeman},
    journal={arXiv},
    year={2024}
}
```