MiniT2I: MSCOCO Captions and Statistics for FID Evaluation

30,000 MSCOCO-2014 validation captions used for FID evaluation in the MiniT2I PyTorch/Diffusers release. The assets include a JSON file of captions and an NPZ file containing reference Inception statistics (mu and sigma arrays) for 512x512 images. The data was created by MiniT2I and last updated on June 14, 2026.

Use Cases

• Compute FID scores for generated images based on the provided reference statistics
• Benchmark text-to-image models using the MSCOCO-30K caption set
• Extract image features for comparison using the pytorch-fid InceptionV3 feature extractor

Strengths

• Includes 30,000 validation captions from the MSCOCO-2014 dataset
• Provides pre-computed reference Inception statistics (mu and sigma arrays) for 512x512 images

Limitations

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

Provenance

Source

MiniT2I

Collection Method

Derived from the MSCOCO-2014 validation set.

Freshness

Last updated 2026-06-14 03:32:20; freshness should be verified