With new image formats such as AVIF (based on AV1), is it worth compressing your images for neural network training to save on disk space? Short answer is no.

AVIF is gaining support across many different softwares and libraries, offering significant disk space savings compared to JPEG (see below). At the same time, image datasets are getting larger and larger, and it could be useful to compress them with smarter algorithms to reduce their size. However, complex compression requires more operations for decoding and may slow down the training loop.

What's the trade-off between image size and decoding time?

To answer this question, I compared the decoding times from disk to tensor of different algorithms. In addition to different PNG settings, I also tested AVIF available to PyTorch through the torchvision-extra-decoders package.

Using the DIV2K validation set, which is composed of 100 high-resolution images (~2000x1000), I measured the average decoding time and file size of several lossless compression algorithms.

The compression algorithms compared are:

• PNG with compression level 9 (maximum compression)
• PNG with compression level 6 (default compression)
• PNG with compression level 0 (no compression)
• AVIF lossless (libaom-av1)
• ZARR, a library for tensor storage (some compression, default parameters)

Experiments are run on my laptop (4 CPU cores, NVMe M.2) and may not represent performance in a cluster environment with more CPUs and shared storage.

While AVIF offers better compression than PNG, its decoding process is much slower. AVIF is too slow for use in a training loop but may be useful for long-term storage or distributing datasets. Uncompressed images (PNG-0) or direct tensor storage (ZARR) produce larger files but are quicker to load into memory.