README.md

---
license: cc-by-nc-4.0
task_categories:
  - image-classification
  - visual-question-answering
tags:
  - image-editing
  - benchmark
  - synthetic-data
  - vlm-as-a-judge
  - croissant
pretty_name: EditJudge-Bench
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: benchmark.parquet
---

# EditJudge-Bench

EditJudge-Bench is a synthetic benchmark for auditing vision-language models used as
automated judges for image-edit verification. Each row contains a source image,
an edited image, a factual edit instruction, counterfactual instructions, and
ground-truth scene parameters produced by a controlled Blender/Infinigen
generation pipeline.

This repository is an anonymous review release for a NeurIPS Evaluations and
Datasets submission.

## Dataset Contents

- 1,500 edit pairs.
- 3,000 JPEG images, stored as paired `before` and `after` files.
- 10 stored edit categories, with 150 examples each.
- The paper reports 9 categories by combining `object_material` and
  `material_slot` into a single material category.
- One row per edit pair, not one row per positive/negative triplet.

Edit-type counts:

- `articulation`: 150
- `camera`: 150
- `lighting`: 150
- `material_slot`: 150
- `movement`: 150
- `object_material`: 150
- `removal`: 150
- `rotation`: 150
- `scale`: 150
- `shape`: 150

## Files

```text
benchmark.parquet
croissant.json
images/<sample_id>/before.jpg
images/<sample_id>/after.jpg
```

The `before` and `after` columns in `benchmark.parquet` are repository-relative
paths. Resolve them relative to the downloaded dataset snapshot.

## Row Format

Important columns include:

- `sample_id`: stable row identifier.
- `edit_type`: edit category.
- `before`, `after`: repository-relative image paths.
- `instruction_pos`: factual instruction that matches the image pair.
- `instruction_neg_list`: list of counterfactual instructions for the same image
  pair.
- `instruction_neg_types`: counterfactual type for each negative instruction.
- `metadata` and `meta.*` / `params.*`: ground-truth procedural parameters saved
  from the Blender generation process.

To evaluate a judge, score the positive triplet
`(before, after, instruction_pos)` and compare it against negative triplets
`(before, after, instruction_neg_list[i])`.

## Intended Use

EditJudge-Bench is intended for diagnostic evaluation of VLM-based editing judges:
whether a model can verify that an image edit was correctly executed, reject
counterfactual edit instructions, and expose failures as a function of known
ground-truth scene parameters.

## Out-of-Scope Use

EditJudge-Bench is not intended as a general-purpose training set for image
generation, a benchmark for human ability, or a safety-critical evaluation
resource. The images are synthetic indoor scenes and should not be treated as
representative of all real-world editing scenarios.

## Responsible AI Notes

The data are rendered synthetic indoor scenes and are not scraped from people,
social media, surveillance footage, or private sources. The benchmark may still
reflect the procedural biases of the scene generator, including object
categories, room layouts, material distributions, and camera placement.

Known limitations include the synthetic-to-real gap, JPEG compression artifacts,
coverage of a finite set of edit types, and the fact that the dataset audits
judges rather than image-editing models directly.

## License

This dataset is released under CC-BY-NC-4.0.

## Croissant

Machine-readable metadata are provided in `croissant.json`, including
Responsible AI metadata for dataset review.

Dataset URL used in metadata: `https://huggingface.co/datasets/EDAnonSubmission/benchmark`.