The Dataset Viewer has been disabled on this dataset.

🌱 PhysGaia: A Physics-Aware Benchmark with Multi-Body Interactions for Dynamic Novel View Synthesis (CVPR 2026)

Mijeong Kim* · Gunhee Kim* · Jungyoon Choi · Wonjae Roh · Bohyung Han

🚀 News

  • [2026-02] PhysGaia has been accepted to CVPR 2026! 🎊
  • [2025-05] Project page and dataset are now available.
  • PhysGaia is a pioneering physics-aware dataset for dynamic novel view synthesis aiming beyond photorealism toward physical realism!

⭐️ Key Highlights

  • 💥 Multi-body interaction
  • 💎 Various materials across all modalities: Liquid, Gas, Viscoelastic substance, and Textile
  • ✏️ Physical evaluation: physics parameters & ground-truth 3D trajectories
  • 😀 Research friendly!!
    • Providing codes for recent Dynamic Novel View Synthesis (DyNVS) models
    • Supporting diverse training setting: both monocular & multiview reconstruction

📂 Dataset Structure


Each folder is corresponding to each scene, containing the following files:
{material_type}_{scene_name}.zip
│
├── render/                              # Generated images (e.g., avatar renderings)
│   ├── train/                           # Images for training
│   └── test/                            # Images for evaluation
│
├── point_cloud.ply                      # COLMAP initialization (PatchMatch & downsampling)
├── camera_info_test.json                # Monocular camera info for test
├── camera_info_train_mono.json          # Monocular camera info for training
├── camera_info_train_multi.json         # Multi-view camera info for training
│
├── {scene_name}.hipnc                   # Houdini source file (simulation or scene setup)
├── particles/                           # Ground-truth trajectories

👩🏻‍💻 Code implementation

Please check each branch for integrated code for recent DyNVS methods.

🔥 Ideal for “Next” Research


  • 🧠 Physical reasoning in dynamic scenes
  • 🤝 Multi-body physical interaction modeling
  • 🧪 Material-specific physics solver integration
  • 🧬 Compatibility with existing DyNVS models

💳 Citation

 @inproceedings{kim2026physgaia,
                            title={PhysGaia:Physics-Aware Benchmark with Multi-Body Interactions for Dynamic Novel View Synthesis},
                            author={Kim, Mijeong and Kim, Gunhee and Choi, Jungyoon and Roh, Wonjae and Han, Bohyung},
                            booktitle={CVPR},
                            year={2026}
                          }                        
                    

🤝 Contributing


We welcome contributions to expand the dataset (additional modality for new downstream tasks, , implementation for other models, etc.) Reach out via opening an issue/discussion in the repo.

🛠️ Future Plans


  • Update fidelity of the generated scenes
  • Add more easier scenes: providing more accessible starting points
  • Add guidelines using Houdini source files: ex) How to obtain a flow field?

💳 License


This project is released under the Creative Commons Attribution-NonCommercial 4.0 license. ✅ Free to use, share, and adapt for non-commercial research

Downloads last month
218

Paper for mijeongkim/PhysGaia