README.md

---
license: cc-by-4.0
task_categories:
- image-to-video
tags:
- simulation
- clevr
- objects
- 3d
- video
- prediction
---

# OBJ3D Dataset

OBJ3D dataset originally from [G-SWM](https://github.com/zhixuan-lin/G-SWM). It contains video frames of synthetic CLEVR-like objects colliding.

This dataset is used for evaluating object-centric world models and stochastic dynamics, as featured in the paper: **[Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling](https://huggingface.co/papers/2603.04553)**.

- **Project Page:** [https://taldatech.github.io/lpwm-web](https://taldatech.github.io/lpwm-web)
- **GitHub Repository:** [https://github.com/taldatech/lpwm](https://github.com/taldatech/lpwm)

## Sample Usage

To train models on this dataset using the official implementation, you can use the following commands:

### Single-GPU Training (DLPv3)
```bash
python train_dlp.py --dataset obj3d_img
```

### Multi-GPU Training (LPWM)
```bash
accelerate launch --config_file ./accel_conf.yml train_lpwm_accelerate.py --dataset obj3d128
```

## Citation

```bibtex
@inproceedings{
daniel2026latent,
title={Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling},
author={Tal Daniel and Carl Qi and Dan Haramati and Amir Zadeh and Chuan Li and Aviv Tamar and Deepak Pathak and David Held},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=lTaPtGiUUc}
}
```