---
license: cc-by-nc-4.0
task_categories:
  - visual-question-answering
  - image-to-text
language:
  - en
tags:
  - vlm
  - benchmark
  - comparative-reasoning
  - subtle-difference
  - image-comparison
  - multi-image
size_categories:
  - 10K<n<100K
---

# VLM-SubtleBench

**VLM-SubtleBench: How Far Are VLMs from Human-Level Subtle Comparative Reasoning?**

The ability to distinguish subtle differences between visually similar images is essential for diverse domains such as industrial anomaly detection, medical imaging, and aerial surveillance. While comparative reasoning benchmarks for vision-language models (VLMs) have recently emerged, they primarily focus on images with large, salient differences and fail to capture the nuanced reasoning required for real-world applications.

VLM-SubtleBench is a benchmark designed to evaluate VLMs on **subtle comparative reasoning** — detecting fine-grained differences between highly similar image pairs that are easy for humans but challenging for state-of-the-art VLMs. Unlike prior benchmarks restricted to natural image datasets, VLM-SubtleBench spans diverse domains including industrial, aerial, and medical imagery.

## Key Findings

- **Proprietary VLMs still struggle** with subtle visual comparison, leaving large gaps from human performance (best model: 77.8% vs. human: 95.5%).
- **Simple prompting strategies** such as chain-of-thought, grid layouts, and overlapping images yield only limited improvements.
- **VLMs are highly sensitive** to difficulty factors such as object size and count, with performance degrading sharply as scene complexity increases.
- **Explicit reasoning helps**: models with stronger inherent reasoning capabilities achieve higher accuracy across all difference types.

## Benchmark Summary

| | |
|---|---|
| **Total QA pairs** | 12,923 |
| **Difference types** | 10 |
| **Image domains** | 6 (Natural, Industrial, Aerial, Synthetic, Medical) |
| **Data sources** | 14 |
| **Human captions** | 1,200 |
| **Splits** | test (11,688) / val (1,235) |
| **Task format** | Multiple-choice VQA + Image Difference Captioning |

> **Note**: Medical domain images (MIMIC-CXR, 362 pairs) are not included due to licensing restrictions, but their QA entries are included in `qa.json`. See [Medical Data](#medical-data-mimic-cxr) below for instructions on how to obtain the images.

## Difference Types

VLM-SubtleBench covers 10 representative difference types, spanning from low-level visual variations to high-level semantic changes:

| Category | Description | Count |
|----------|-------------|-------|
| Attribute | Variations in object properties (color, size, shape) | 1,196 |
| State | Object condition changes (breakage, cracks, peeling) | 1,148 |
| Emotion | Comparative judgments of facial expression intensity | 1,108 |
| Temporal | Identifying which image depicts an earlier/later event | 1,117 |
| Spatial | Changes in arrangement or relative position | 1,235 |
| Existence | Whether an object has appeared or disappeared | 1,204 |
| Quantity | Changes in object count | 1,599 |
| Quality | Degradations such as blur, noise, or overexposure | 1,109 |
| Viewpoint | Camera perspective changes (pan, tilt, roll, orbit) | 1,734 |
| Action | Differences in human/object poses or activities | 1,111 |

## Image Domains and Sources

| Domain | Count | Sources |
|--------|-------|---------|
| Natural | 7,526 | CameraBench, ChangeIt, COCO, CREMA-D/RAVDESS/AFEW-VA/DAiSEE, MegaFruits, UCF-QNRF, VLM4D, YouTube-8M |
| Synthetic | 3,190 | Procedurally generated primitive scenes (circles, squares, triangles on white backgrounds) |
| Industrial | 1,202 | MVTec-AD, MVTec-LOCO |
| Aerial | 643 | LEVIR-MCI, UBC |
| Medical | 362* | MIMIC-CXR *(not included, see below)* |

## Dataset Structure

```
VLM-SubtleBench/
├── README.md
├── qa.json                          # All QA pairs with metadata (including MIMIC)
└── images/
    ├── camerabench/                 # CameraBench — viewpoint
    ├── changeit/                    # ChangeIt — state, quantity
    │   ├── state/
    │   └── quantity/
    ├── coco/                        # COCO + Gemini edits — attribute, existence
    │   ├── val2017/
    │   ├── val2017_attribute_edit/
    │   ├── val2017_existence_edit/
    │   ├── train2017/
    │   ├── train2017_attribute_edit/
    │   └── train2017_existence_edit/
    ├── emotion_videos/              # CREMA-D/RAVDESS/AFEW-VA/DAiSEE — emotion
    ├── levir/                       # LEVIR-MCI — existence, quantity
    │   ├── existence/
    │   └── quantity/
    ├── megafruits/                   # MegaFruits — quantity
    ├── mvtec_ad/                    # MVTec-AD — attribute, state
    ├── mvtec_loco/                  # MVTec-LOCO — quantity
    ├── synthetic/                   # Synthetic primitives — attribute, existence, quantity, spatial, viewpoint
    │   ├── attribute/
    │   ├── existence/
    │   ├── quantity/
    │   ├── spatial/
    │   └── viewpoint/
    ├── ubc/                         # UBC — quantity
    ├── ucf_qnrf/                    # UCF-QNRF — quantity
    ├── vlm4d/                       # VLM4D — spatial, temporal
    │   ├── spatial/
    │   └── temporal/
    └── yt8m/                        # YouTube-8M — action, quality, temporal
        ├── action/
        ├── quality/
        └── temporal/
```

## QA Entry Format

Each entry in `qa.json`:

```json
{
  "image_1": "images/camerabench/processed_frames/1018.1.7/frame_1.png",
  "image_2": "images/camerabench/processed_frames/1018.1.7/frame_2.png",
  "question": "In which direction does the camera move from the first image to the second image?",
  "answer": "backward",
  "distractors": ["forward"],
  "has_caption": false,
  "caption": null,
  "split": "test",
  "metadata": {
    "category": "viewpoint",
    "domain": "natural",
    "source": "camerabench",
    "source_id": "0",
    "raw_folder": "camera_pairs",
    "generation_info": {
      "movement_type": "dolly-out",
      "original_labels": ["minimal-shaking", "complex-motion", "regular-speed", "dolly-out", "lead-tracking"],
      "video_path": "videos_gif/1018.1.7.gif"
    }
  }
}
```

| Field | Description |
|-------|-------------|
| `image_1`, `image_2` | Relative paths to the image pair |
| `question` | The comparative question about the two images |
| `answer` | Correct answer |
| `distractors` | Incorrect answer choices |
| `has_caption` | Whether a human-written difference caption is available |
| `caption` | Human-written description of the difference between the images (null if unavailable) |
| `split` | `test` or `val` |
| `metadata.category` | One of the 10 difference types |
| `metadata.domain` | Image domain (natural, industrial, aerial, synthetic, medical) |
| `metadata.source` | Source dataset identifier |
| `metadata.source_id` | Original ID within the source dataset |
| `metadata.generation_info` | Source-specific metadata (varies by source, may be null) |

## Medical Data (MIMIC-CXR)

The medical domain QA entries (362 attribute comparison pairs from MIMIC-CXR chest X-rays) are included in `qa.json`, but the corresponding images are not included due to [PhysioNet licensing requirements](https://physionet.org/content/mimic-cxr-jpg/2.1.0/).

To obtain the medical images:

1. Obtain credentialed access to [MIMIC-CXR-JPG v2.1.0](https://physionet.org/content/mimic-cxr-jpg/2.1.0/) on PhysioNet
2. Download the required chest X-ray images
3. Place them under `images/mimic/` following the path structure referenced in `qa.json` (e.g., `images/mimic/p15/p15592981/s55194630/{hash}.jpg`)

The image paths preserve the original MIMIC-CXR directory hierarchy, so files can be copied directly from a standard MIMIC-CXR-JPG download.

## Loading and Evaluation

To be added.

## Citation

To be added.