AIPeanutman/

OSBench

OSBench is a subject-centric benchmark for subject-driven image generation and manipulation. OSBench explicitly covers both generation and manipulation under more challenging settings.

Total: 4 sub-tasks × 150 samples = 600 samples.
Usage: evaluation-only benchmark (not intended for large-scale training).

We define four sub-tasks (each with 150 samples):

Single-subject generationSynthesize an identity-consistent image from one reference under an… See the full description on the dataset page: https://huggingface.co/datasets/AIPeanutman/OSBench.