---
pretty_name: AmbiSem-3D
tags:
- 3d
- image
- amodal-3d-generation
- semantic-ambiguity
---

# AmbiSem-3D

AmbiSem-3D is a diagnostic benchmark for amodal 3D generation under semantic
ambiguity, introduced with [**RelaxFlow: Text-Driven Amodal 3D
Generation**](https://arxiv.org/abs/2603.05425), an ICML 2026 Spotlight paper.

This repository contains two separated subsets:

| Subset | Folder | Entries | Description |
|---|---:|---:|---|
| Original | `original/` | 21 | Hand-curated single-image ambiguity cases |
| Extended | `extended/` | 100 | Semi-automatically curated view-induced ambiguity cases |

Each subset has its own `manifest.json`, `README.md`, and `assets/` directory.
All paths in a subset manifest are relative to that subset folder.
Code and runners are available in the
[**RelaxFlow repository**](https://github.com/viridityzhu/RelaxFlow).

## Usage

Download the repository, then pass the subset manifest path to the RelaxFlow
batch runner or to the `prepare_manifest_with_priors.py` helper from the
RelaxFlow code release.

## Citation

If you use AmbiSem-3D or RelaxFlow, please cite:

```bibtex
@inproceedings{zhu2026relaxflow,
  title     = {RelaxFlow: Text-Driven Amodal 3D Generation},
  author    = {Zhu, Jiayin and Fu, Guoji and Liu, Xiaolu and He, Qiyuan and Li, Yicong and Yao, Angela},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026},
  url       = {https://arxiv.org/abs/2603.05425}
}
```