CelebA-C: Noise Corruptions

This dataset is part of the CelebA-C benchmark, a large-scale image corruption dataset derived from the CelebA face dataset for evaluating model robustness under distribution shift.

This record contains the noise-based corruption subset, including

Gaussian noise,

shot noise,

impulse noise, and

speckle noise

each generated at five severity levels on the CelebA test split.

Images are stored in HDF5 format with lossless gzip compression and chunked storage to enable efficient loading and reproducible evaluation.

The complete CelebA-C benchmark is released as four complementary datasets due to repository size constraints:(i) Celeba_blur.h5(ii) Celeba_noise.h5(iii) Celeba_digital.h5, and(iv) Celeba_weather.h5 (refer to the DOIs of the other three datasets in the Related Works section)Together, these four records constitute the complete CelebA-C dataset.