README.md

---
pretty_name: SpatialGen-Bench
language:
- en
size_categories:
- n<1K
task_categories:
- visual-question-answering
- image-to-image
license: other
tags:
- spatial-intelligence
- spatial-reasoning
- image-generation
- vision-language-models
- benchmark
dataset_info:
- config_name: default
  features:
  - name: id
    dtype: string
  - name: capability
    dtype: string
  - name: task
    dtype: string
  - name: image_path
    dtype: string
  - name: media_paths
    sequence: string
  - name: mask_path
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  - name: answer_type
    dtype: string
  - name: choices
    sequence: string
  - name: source
    dtype: string
  - name: source_id
    dtype: string
  - name: metadata_json
    dtype: string
  splits:
  - name: test
    num_examples: 470
configs:
- config_name: default
  data_files:
  - split: test
    path: data/spatialgen_bench.jsonl
---

<p align="center">
  <video controls playsinline preload="metadata" poster="https://huggingface.co/datasets/wx91726/SpatialGen-Bench/resolve/main/assets/provise-overview-poster.webp" width="90%">
    <source src="https://huggingface.co/datasets/wx91726/SpatialGen-Bench/resolve/main/assets/provise-overview.mp4" type="video/mp4">
  </video>
</p>

<p align="center">
  <a href="https://arxiv.org/abs/2607.21072"><img src="https://img.shields.io/badge/arXiv-2607.21072-B31B1B?style=flat-square" alt="Paper"></a>
  <a href="https://zju-omniai.github.io/ProVisE/"><img src="https://img.shields.io/badge/Project-Page-1769AA?style=flat-square" alt="Project"></a>
  <a href="https://github.com/ZJU-OmniAI/ProVisE"><img src="https://img.shields.io/badge/Code-ProVisE-181717?style=flat-square&amp;logo=github&amp;logoColor=white" alt="Code"></a>
</p>

## Benchmark

<p align="center">
  <img src="assets/benchmark_overview.webp" alt="SpatialGen-Bench task overview across perception, understanding, reasoning, and interaction." width="100%">
</p>

Each record retains its source-task metric target as `integer`, `text`, `point`, `mask`, or `polyline`. The frozen visual-answer contract is available at [`protocols/spatialgen_bench.yaml`](protocols/spatialgen_bench.yaml), with runtime parsers in [ProVisE](https://github.com/ZJU-OmniAI/ProVisE).

## Quick Start

```python
from datasets import load_dataset

dataset = load_dataset("wx91726/SpatialGen-Bench", split="test")
print(dataset[0])
```

Download the complete media and evaluation package for local evaluation:

```bash
hf download wx91726/SpatialGen-Bench \
  --repo-type dataset \
  --local-dir SpatialGen-Bench
```

<details>
<summary><strong>Dataset structure and JSONL schema</strong></summary>

```text
SpatialGen-Bench/
|-- data/spatialgen_bench.jsonl
|-- benchmarks/{perception,understanding,reasoning,interaction}/
|-- protocols/spatialgen_bench.yaml
|-- SOURCES.md
`-- LICENSE
```

- `id`: stable sample identifier
- `capability`: one of the four capability levels
- `task`: subtask name
- `image_path`: primary input path relative to `benchmarks/`
- `media_paths`: ordered inputs for single- or multi-image tasks
- `mask_path`: optional ground-truth mask path
- `question`, normalized `answer`, `answer_type`, and `choices`
- `source` and `source_id`: upstream provenance
- `metadata_json`: serialized task-specific evaluator metadata

</details>

## Sources and License

SpatialGen-Bench annotations and protocol configuration are released under Apache-2.0. Source-derived media remains subject to the licenses and terms of its upstream datasets. [`SOURCES.md`](SOURCES.md) records task-level provenance, source links, and usage terms.

<details>
<summary><strong>Acknowledgements and upstream resources</strong></summary>

We sincerely appreciate [CountBench](https://teaching-clip-to-count.github.io/), [BLINK](https://huggingface.co/datasets/BLINK-Benchmark/BLINK), [EgoOrientBench](https://huggingface.co/datasets/jhCOR/EgoOrientBench), [VSR](https://github.com/cambridgeltl/visual-spatial-reasoning), [ViewSpatial-Bench](https://huggingface.co/datasets/lidingm/ViewSpatial-Bench), [MindCube](https://huggingface.co/datasets/MLL-Lab/MindCube), [VisWorld-Eval](https://github.com/thuml/Reasoning-Visual-World), [RoboAfford-Eval](https://huggingface.co/datasets/tyb197/RoboAfford-Eval), [ShareRobot-Bench](https://huggingface.co/datasets/BAAI/ShareRobot-Bench), [PhysBench](https://huggingface.co/datasets/USC-PSI-Lab/PhysBench), [SPHERE-VLM](https://sphere-vlm.github.io/), and [RefCOCOg](https://github.com/lichengunc/refer) for their public datasets, task designs, and evaluation resources.

</details>

## Citation

```bibtex
@article{wang2026showdonttell,
  title  = {Show, Don't Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text},
  author = {Wang, Xu and Yao, Kaixiang and Pan, Miao and Zhou, Xiaohe and Liu, Xuanyu and Zhang, Wenqi and Zhang, Xuhong},
  journal = {arXiv preprint arXiv:2607.21072},
  year   = {2026},
  url    = {https://arxiv.org/abs/2607.21072}
}
```