---
license: cc-by-4.0
pretty_name: GenScale
task_categories:
- text-to-image
- image-to-image
- image-classification
language:
- en
size_categories:
- 1K<n<10K
configs:
- config_name: entries
  data_files:
  - split: train
    path: data/genscale_entries.parquet
- config_name: pairs
  data_files:
  - split: train
    path: data/genscale_pairs.parquet
- config_name: task3_edit_plans
  data_files:
  - split: train
    path: data/genscale_task3_edit_plans.parquet
---

# GenScale

**GenScale: A Benchmark for Relative Object Scale in Image Generation and Editing**  
Lingxiao Li, Max Whitton, Ledell Wu, and Boqing Gong.

[Project page](https://lingxiao-li.github.io/genscale.github.io/) · [Paper](https://arxiv.org/abs/2609.00525) · [Code](https://github.com/lingxiao-li/GenScale) · [Rescale checkpoint](https://huggingface.co/lingxiao2049/Rescale)

## Contents and loading

| Configuration | Rows | Meaning |
|---|---:|---|
| `entries` | 900 | Generation/editing prompts and metadata |
| `pairs` | 1,643 | Pairwise scale evaluation units |
| `task3_edit_plans` | 100 | Precise S5 editing instructions |

```python
from datasets import load_dataset
entries = load_dataset("lingxiao2049/GenScale", "entries", split="train")
pairs = load_dataset("lingxiao2049/GenScale", "pairs", split="train")
```

The `train` split name is a storage convention, not a prescribed training split. CSV mirrors are provided alongside Parquet. For images and the complete JSON, use `huggingface_hub.snapshot_download(repo_id="lingxiao2049/GenScale", repo_type="dataset", local_dir="GenScale")`.

`GenScale_Benchmark.json` is the canonical JSON entry point. It is byte-identical to the retained `GenScale_Benchmark_anonymous_clean.json` from the anonymous review release. Scenario counts are S1=200, S2=200, S3=300, S4=100, S5=100 entries; corresponding pair counts are 570, 573, 300, 100, 100. S4 and S5 share 100 source images. The 433 unique included images comprise 294 Task 2 references, 100 Task 3 inputs, and 39 Task 3 references. Every nonempty input/reference path in the benchmark JSON resolves within this release.

The two `metadata/` CSVs contain the 100-category physical-size knowledge base and the 1,000-product ABO candidate pool. Candidate-pool `abo_images/original/...` paths identify upstream archive assets, which are not bundled; benchmark inputs use the included `images/...` paths. `scripts/make_genscale_parquet.py` regenerates tables from the frozen JSON. `SHA256SUMS` records checksums; `VALIDATION.json` records release checks.

## Scale convention

`object_a`, `object_b`, and `target_ratio_a_to_b` correspond to the key `object_a_to_object_b`: the intended physical characteristic-length ratio **length(A) / length(B)**. `acceptable_ratio_min/max` use the same direction. The judge code uses A as anchor, B as target, and estimates **B/A**; take the reciprocal when converting between them. Physical-scale metadata and generated prose can disagree on individual entries; preserve and inspect both rather than inferring a new ground truth from prompt text.

## Collection, processing, and responsible use

This benchmark combines a common-object physical-size knowledge base, ABO-derived product dimensions/catalog reference images, API-assisted image-generation prompts, and selected generated inputs for scale editing. The frozen JSON and conversion script define the released records; `PROCESSING_REPORT.md` preserves the original release counts. Local absolute paths were replaced with relative asset paths. Missing scenario-specific fields are represented as null/empty cells and have not been imputed. Ratios and annotations are preserved from the review release.

The intended use is research evaluation of relative object scale in image generation and editing. The benchmark is small, uses English prompts and selected everyday objects/products, and is not representative of all cultures, bodies, physical environments, or object categories. Metadata dimensions may be imperfect, prompt text can conflict with metadata, and generated inputs can contain artifacts. A plausible scale score does not establish visual quality, factual correctness, or suitability for safety-critical measurement. Human-product prompts use simplified fixed body-size anchors; these are evaluation assumptions, not population statistics.

The public release contains prompts, object/product metadata, and images. It does not include private human annotation forms, contact details, or rater identities. Product imagery and synthesized people may depict faces or brands; no biometric labels or person-identification ground truth are supplied. No independent privacy, demographic fairness, or harmful-content audit is claimed. Human-study compensation and ethics-review status are outside this dataset release and are not inferred here.

Machine-assisted construction metadata records Gemini model identifiers where available. The paper's scoring pipeline is provided in the code repository; benchmark rows are not themselves the complete per-model judge-output archive. Trained Rescale checkpoints are distributed separately at [lingxiao2049/Rescale](https://huggingface.co/lingxiao2049/Rescale) under the FLUX non-commercial license. This dataset does not bundle those weights, all generated model outputs, or raw human-study responses. Report bugs through the linked GitHub repository; cite a Hugging Face commit revision for reproducible use.

[`croissant.json`](croissant.json) is the canonical, author-maintained Croissant metadata file for all three structured tables. It includes all eight minimum RAI topics in the [NeurIPS 2026 hosting guidance](https://neurips.cc/Conferences/2026/EvaluationsDatasetsHosting): limitations, biases, personal/sensitive information, intended uses and construct validity, social impact, a synthetic-data boolean, source datasets, and collection/preprocessing/annotation provenance. Synthetic source prompts and available generator identifiers are documented; missing generation seeds and backend revisions are explicitly identified. CSV sources are pinned to an immutable public dataset revision and have SHA-256 checksums.

The exact file passed all four checks in the [conference-linked Croissant checker](https://huggingface.co/spaces/JoaquinVanschoren/croissant-checker) (JSON, schema, RAI fields, and record generation). See [CROISSANT_VALIDATION_REPORT.txt](CROISSANT_VALIDATION_REPORT.txt) and [RAI_REQUIREMENTS.json](RAI_REQUIREMENTS.json). The checker currently tests six RAI keys; the two required PROV source/activity fields were additionally reviewed against the conference guidance. Local `mlcroissant` validation loaded all 900 / 1,643 / 100 records. **For OpenReview, upload this repository's `croissant.json`, which contains the authored RAI information; an automatically generated metadata export may omit these additions.**

## License and attribution

Dataset: **CC BY 4.0**, preserving the anonymous release's license. See [LICENSE](LICENSE) and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) for ABO attribution and external-asset scope. Code is separately MIT licensed. Model weights retain their upstream terms.

Please cite [CITATION.bib](CITATION.bib).