Subjects2K: A 2,000-Pair Benchmark for Identity Preservation in Generated Images

Subjects2K is a benchmark dataset of 2,000 image pairs designed for evaluating identity preservation in generated or edited images. It was created by chaenayo to evaluate the ID-Sim metric introduced in a CVPR 2026 paper. The dataset is a stratified subset derived from the larger Subjects200K collection.

Use Cases

• Benchmarking identity preservation metrics based on the described gpt_score stratification.
• Evaluating face-swapping or image-editing models on identity consistency.
• Training or validating models for identity-aligned similarity measurement.
• Researching the correlation between human perception of identity and automated scores.

Strengths

• Contains exactly 2,000 image pairs, providing a defined benchmark scale.
• Pairs were selected via stratified sampling over the gpt_score from a larger source, suggesting controlled diversity.
• Created for a specific, peer-reviewed research purpose (evaluating ID-Sim at CVPR 2026).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the source data is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Derived from the Subjects200K dataset.

Collection Method

2,000 pairs were selected by stratifying over the original gpt_score in Subjects200K.

Freshness

Last updated 2026-06-01 06:51:13; freshness should be verified.