README.md

metadata

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

Benchmark

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, with runtime parsers in ProVisE.

Quick Start

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:

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

Dataset structure and JSONL schema

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

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 records task-level provenance, source links, and usage terms.

Acknowledgements and upstream resources

We sincerely appreciate CountBench, BLINK, EgoOrientBench, VSR, ViewSpatial-Bench, MindCube, VisWorld-Eval, RoboAfford-Eval, ShareRobot-Bench, PhysBench, SPHERE-VLM, and RefCOCOg for their public datasets, task designs, and evaluation resources.

Citation

@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}
}