Testing Google’s Guetzli JPEG Compression: Performance, Memory Usage, and Practical Tips

Google recently released a new JPEG compression algorithm called Guetzli, which aims to retain details that the human eye can easily recognize while discarding those that are less noticeable, resulting in smaller files with comparable perceived quality.

In a real‑world project with a homepage of about 30 MB (27 MB of which were images), the author tested Guetzli on a high‑resolution 8574×5715 JPEG (22 MB). The first attempt failed due to insufficient memory; Guetzli requires roughly 300 MB of RAM per megapixel, meaning the test image needed about 15 GB.

After freeing up memory, the second run succeeded, using 12 GB of RAM and taking a little over 40 minutes (approximately one minute per megapixel as claimed by Google). The compressed image was under 7 MB, and visual differences were described as “imperceptibly small.”

Using Guetzli’s default quality range (84–100), the author experimented with various quality factors. The results showed a clear trade‑off: lower quality values produced much smaller files (e.g., 3.0 MB at quality 84) while higher values yielded larger files (up to 18 MB at quality 99). The product owner and designer agreed to use quality 84, reducing the homepage size from 30 MB to under 8 MB.

Installation of Guetzli is straightforward on Linux (via an AUR package) and also available as a Homebrew package for macOS. The tool requires ample RAM and CPU time; it runs single‑threaded, so multiple instances can be parallelized if memory permits.

Guetzli only outputs JPEG files (no PNG or transparency) but can compress PNGs by converting them to JPEG. Compression ratios vary with the original image size and quality, typically 2×–7× reduction, with more noticeable quality loss on smaller images.

In summary, Guetzli can significantly reduce image sizes for web pages, but users must be prepared for high memory consumption and long processing times, especially for large images. Choosing an appropriate quality setting (around 84) offers a good balance between size reduction and visual fidelity.