Guetzli

Guetzli is known to be very resource intensive, requiring several orders of magnitude more processing time than other JPEG encoders, and a lot of memory.[4] Guetzli supports only the top of JPEG's quality range (quantizer settings 84–100)[5][6] and supports only sequential (non-"progressive") encoding. The performance of the Butteraugli quality metric it optimizes for breaks down below that range and starves some areas too much. Guetzli is more effective with bigger files.[6]

It is often placed in line with other bits of Google technology that are used to cut down web page load times, while Google itself describes it also as a demonstration of the potential of psychovisual optimizations to motivate further research to benefit future JPEG encoders,[2] and acknowledges the often impractically low speed of their code.[7]

Two tests found that Guetzli is very slow (about 4 magnitudes slower than normal JPEG encoder) and not necessarily better than mozjpeg.[5][8]

It is implemented as a command-line tool in C++[9] and the source code published on GitHub as free software under the terms of version 2 of the Apache license. Several important Linux distributions have packages available in their official repositories.[10][11][12] For Arch Linux there are user repositories available.[13][14] There is a macOS version packaged in the Homebrew repository.[15] There are graphical user interfaces and prebuilt Windows binaries available elsewhere on the Internet.[4][16][17] For web development with node.js, there is a Guetzli package available in the Node Package Manager (npm).[18]

The first public version was released on October 21, 2016, without any speed optimizations,[15] and only announced on a specialist forum.[19][20] Version 1.0 followed five months later on March 15, 2017,[21] accompanied by an announcement to a broader public and two scientific papers.[22][23][24]

Guetzli uses methods to optimize compression efficiency that target mainly the quantization step. It constructs custom quantization tables for each file, decides on color subsampling,[25] and quantizes some adjacent DCT coefficients to zero, balancing benefits in the run-length encoding of coefficients and preservation of perceived image fidelity. Zeroing the right coefficients is the most effective tool in Guetzli, which is used as a makeshift means of spatially adaptive quantization. The optimizations are guided by Butteraugli.[26]

Butteraugli is a newly introduced objective image quality assessment metric. It assigns a differential mean opinion score (DMOS) value to the difference between an original image and a degraded version. It is significantly more complex than traditional metrics like PSNR and SSIM, but claimed to perform better with high-end quality, where degradations are not or barely noticeable.[27] It models color perception and visual masking in the human visual system, taking into account that the eye is imaging different colors with different precision.[28] It uses a heat map of changes.[26] How the hundreds of parameters that model the properties of the human visual system were derived remains unexplained.[1] An in-house performance evaluation with 614 ratings from 23 people on their own test set of 31 images yielded 75% of ratings favouring of JPEGs encoded for Butteraugli scores over libjpeg-turbo encodes,[23] which usually score higher on SSIM and PSNRHVS-M.[20][29]

Translating to "butter eye", the Swiss-German name originally signifies a dimple on top of some sweet pastry that has been filled with butter and sugar before baking.[30]

• Guetzli on GitHub
• Butteraugli on GitHub