guetzli

Perceptual JPEG encoder that optimizes for visual quality at small file sizes

TLDR

Re-encode a JPEG

$ guetzli [input.jpg] [output.jpg]

Encode a PNG to JPEG

$ guetzli [input.png] [output.jpg]

Set quality level (84-100)

$ guetzli --quality [90] [input.jpg] [output.jpg]

Maximum compression (lowest supported quality)

$ guetzli --quality [84] [input.jpg] [output.jpg]

Show progress trace

$ guetzli --verbose [input.png] [output.jpg]

SYNOPSIS

guetzli [--quality Q] [--verbose] input output.jpg

Guetzli is a JPEG encoder that targets the human visual system. It uses Butteraugli, a psychovisual model, to search the JPEG quantization space and produce baseline-compatible JPEG files that are typically 20-30% smaller than equivalent libjpeg output at the same perceived quality.Output files are compatible with all standard JPEG decoders. The trade-off is encoding cost: guetzli is computationally and memory intensive.

PARAMETERS

input

Source image. PNG and JPEG inputs are supported.

output.jpg

Destination JPEG file.

--quality Q

Visual quality, expressed in libjpeg-equivalent units. Range 84-100, default 95. Values below 84 are not supported.

--verbose

Print a trace of encoding attempts.

CAVEATS

Very slow: encoding a single megapixel can take a minute or more. Memory usage is roughly 300 MB per megapixel. Only quality values 84-100 are accepted. Best results come from non-degraded sources (PNG or high-quality JPEG); re-encoding heavily compressed JPEGs offers limited gains.

HISTORY

Guetzli was developed by Google Research and released as open source in March 2017 alongside the Butteraugli perceptual distance metric.