OBJ3D Dataset

OBJ3D dataset originally from 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.

• Project Page: https://taldatech.github.io/lpwm-web
• GitHub Repository: 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)

python train_dlp.py --dataset obj3d_img

Multi-GPU Training (LPWM)

accelerate launch --config_file ./accel_conf.yml train_lpwm_accelerate.py --dataset obj3d128

Citation

@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}
}