COCO-2014 Val 10K: Image-Caption Benchmark for FID Evaluation

A curated subset of the COCO 2014 Validation set containing 9,986 image-caption pairs. The images are center-cropped and resized to 256x256 resolution in PNG format. The dataset was created by user 'byliutao' and is designed as a standard reference benchmark for Fréchet Inception Distance evaluation in text-to-image generation research.

Use Cases

• Benchmarking text-to-image model performance based on the FID metric.
• Training or fine-tuning image captioning models using the provided image-caption pairs.
• Conducting research on multimodal representation learning using the standardized image set.

Strengths

• Contains 9,986 standardized image-caption pairs, providing a substantial sample size.
• Images are uniformly processed to 256x256 resolution, ensuring consistency for evaluation.
• Uses a fixed random seed (42) for subset selection, enabling reproducible benchmarking.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is derived from a 2014 source, which may not reflect current visual or linguistic trends.
• Metadata completeness is limited; actual data quality requires manual inspection after download.

Provenance

Source

COCO 2014 Validation Split

Collection Method

Curated subset selected with a fixed random seed from an original pool of 40,504 samples.

Freshness

Last updated 2026-05-11 03:52:48; freshness should be verified.