# Data Sources

SpatialGen-Bench is a curated evaluation split derived from public spatial benchmarks. The annotations and protocol configuration are original SpatialGen-Bench contributions; source media retains its upstream license and terms.

| Source benchmark | SpatialGen-Bench task | Samples | License / terms | Official resources |
|---|---|---:|---|---|
| CountBench | Counting | 40 | CC-BY-4.0; source-image rights retained | [Project](https://teaching-clip-to-count.github.io/) / [HF mirror](https://huggingface.co/datasets/nielsr/countbench) |
| BLINK | Depth | 30 | Apache-2.0 | [HF](https://huggingface.co/datasets/BLINK-Benchmark/BLINK) |
| EgoOrientBench | Orientation | 40 | MIT | [HF](https://huggingface.co/datasets/jhCOR/EgoOrientBench) |
| VSR | Relationship | 35 | Apache-2.0; COCO media terms | [GitHub](https://github.com/cambridgeltl/visual-spatial-reasoning) |
| ViewSpatial-Bench | Perspective | 35 | Apache-2.0; COCO / ScanNet media terms | [HF](https://huggingface.co/datasets/lidingm/ViewSpatial-Bench) |
| MindCube | Mental Modeling | 35 | MIT | [HF](https://huggingface.co/datasets/MLL-Lab/MindCube) |
| VisWorld-Eval | Multi-hop, Prediction | 55 | N/A | [GitHub](https://github.com/thuml/Reasoning-Visual-World) / [HF](https://huggingface.co/datasets/thuml/VisWorld-Eval) |
| RoboAfford-Eval | Affordance | 35 | CC-BY-4.0 | [HF](https://huggingface.co/datasets/tyb197/RoboAfford-Eval) |
| ShareRobot-Bench | Trajectory | 30 | Apache-2.0 | [HF](https://huggingface.co/datasets/BAAI/ShareRobot-Bench) |
| PhysBench | Navigation | 30 | Apache-2.0; source-media terms | [HF](https://huggingface.co/datasets/USC-PSI-Lab/PhysBench) |
| SPHERE-VLM | Object Size, Geometric Feasibility | 70 | N/A; COCO media terms | [Project](https://sphere-vlm.github.io/) / [HF](https://huggingface.co/datasets/wei2912/SPHERE-VLM) |
| RefCOCOg | Spatial Grounding | 35 | Apache-2.0 repository; COCO media terms | [GitHub](https://github.com/lichengunc/refer) / [HF](https://huggingface.co/datasets/lmms-lab/RefCOCOg) |

Licenses and usage terms are recorded from the checked official project page, repository, or dataset card. They document provenance and do not override terms attached to upstream media. `N/A` means that no explicit dataset license was identified in the checked source; it does not imply an absence of restrictions.