Update dataset card with task category, paper link, and usage instructions

README.md

@@ -1,5 +1,7 @@
 ---
 license: cc-by-4.0
+task_categories:
+- image-to-video
 tags:
 - simulation
 - clevr
@@ -8,5 +10,39 @@ tags:
 - video
 - prediction
 ---
-OBJ3D dataset from G-SWM (https://github.com/zhixuan-lin/G-SWM).
-Contains video frames of synthetic CLEVR-like objects colliding.
+
+# 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}
+}
+```