---
task_categories:
- image-to-text
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image_1
    dtype: image
  - name: image_2
    dtype: image
  - name: choices
    struct:
    - name: A
      dtype: string
    - name: B
      dtype: string
    - name: C
      dtype: string
    - name: D
      dtype: string
  - name: ground_truth
    dtype: string
  - name: category
    dtype: string
  splits:
  - name: train
    num_bytes: 4188405701.612
    num_examples: 1756
  download_size: 5002312163
  dataset_size: 4188405701.612
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# VDiff-Bench

[Paper](https://huggingface.co/papers/2609.06245) | [Project page](https://huggingface.co/spaces/elaine1wan/vdiff-bench)

VDiff-Bench is a multiple-choice benchmark for fine-grained visual difference identification. Each example presents two similar images and four candidate descriptions, exactly one of which states a real difference between the images.

## Dataset structure

The `train` split contains 1,756 questions with the following fields:

- `id`: stable example identifier.
- `image_1`, `image_2`: the paired images.
- `choices`: an object containing answer choices `A`, `B`, `C`, and `D`.
- `ground_truth`: label of the correct choice.
- `category`: visual change category.

The public release intentionally excludes source identifiers, intermediate difference annotations, generated-negative metadata, and other dataset-construction fields.

## Change categories

| Category | Questions |
|---|---:|
| Change of position | 158 |
| Change of motion | 138 |
| Change of color (regional) | 170 |
| Change of color (whole image) | 147 |
| Appear / disappear | 201 |
| Change of noise / resolution | 150 |
| Change of texture | 177 |
| Substitution and size | 199 |
| OCR | 221 |
| Illumination | 195 |

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("elaine1wan/image_diff_data", split="train")
example = dataset[0]
```

`image_1` and `image_2` are decoded as PIL images by default. Benchmark performance is measured using exact-match accuracy on `ground_truth`.

## Citation

Citation information will be added with the accompanying paper release.