SpotDiff is a visual spot-the-difference benchmark. Each item is one composite
image containing two nearly identical panels. A model receives the complete
image and identifies every visible difference using structured JSON.
This is the public development release. It contains 10 reviewed images and 61
gold differences. The annotations are intentionally public so that anyone can
reproduce the evaluator locally and inspect the benchmark design.… See the full description on the dataset page:
https://huggingface.co/datasets/Abnik/spotdiff-v1-dev.
