README.md

metadata

language:
  - en
license: apache-2.0
size_categories:
  - 10K<n<100K
pretty_name: st3d
task_categories:
  - text-to-image
tags:
  - 3D-layout
  - controllable-generation
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl

SeeThrough3D Dataset

Project Page | Paper | GitHub

This is the training dataset for the CVPR 2026 🎉 paper SeeThrough3D: Occlusion Aware 3D-Control in Text-to-Image Generation.

SeeThrough3D is a model for 3D layout-conditioned generation that explicitly models occlusions. This dataset consists of diverse multi-object scenes with strong inter-object occlusions, using an occlusion-aware 3D scene representation (OSCR) where objects are depicted as translucent 3D boxes.

Dataset Information

The primary training data is contained in train.jsonl.

The training code expects shuffled versions of the jsonls (train_shuffled{0..3}.jsonl). These files are shuffled versions of train.jsonl with no additional content.

For detailed instructions on environment setup and training, please refer to the official GitHub repository.

Citation

If you find this work or dataset useful, please cite:

@misc{agrawal2026seethrough3docclusionaware3d,
      title={SeeThrough3D: Occlusion Aware 3D Control in Text-to-Image Generation}, 
      author={Vaibhav Agrawal and Rishubh Parihar and Pradhaan Bhat and Ravi Kiran Sarvadevabhatla and R. Venkatesh Babu},
      year={2026},
      eprint={2602.23359},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.23359}, 
}