Dataset · KRAFTON
VLM fine-grained visual comparison benchmark
KRAFTON/VLM-SubtleBench
A collection of images and English questions to test VLM ability to find subtle differences in similar pictures.
What's inside
The dataset from KRAFTON includes between 10k and 100k records with images and visual question answering (VQA) in English. Unlike standard tests with obvious differences, it features visually similar frames with barely noticeable distinctions. The task is designed to evaluate fine comparative analysis skills in vision-language models.
What you can do with it
- Testing VLM ability to find minor discrepancies between two nearly identical images
- Evaluating models for industrial anomaly detection, medical imaging, and aerial photography tasks
- Tasks
- Visual question answering, Image → text
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 29.9K
- Likes
- 7
How to load it
from datasets import load_dataset
ds = load_dataset("KRAFTON/VLM-SubtleBench")