• libjxl 0.10 Release: The reference implementation for JPEG XL, version 0.10, has been released. It fully implements the "streaming encoding" API, allowing large images to be encoded in chunks, reducing memory usage and speeding up encoding, especially for lossless compression of large images.
• Memory and Time Savings: Compressing a 13500×6750 NASA image from 64 MB as a TIFF file to 33.2 MB jxl file now requires only 0.7 gigabytes of RAM and 30 seconds using a single thread (or 5 seconds with eight threads) in libjxl 0.10, compared to 8 gigabytes of RAM and over two minutes in version 0.9.
• Pareto-Optimal Results: The new version of libjxl beats the previous version by a significant margin and achieves Pareto-optimal results across all speeds, producing smaller files than PNG and lossless AVIF or lossless WebP. It is also more Pareto-efficient than other codecs for different types of images.
• Objective Metrics: Objective metrics like SSIMULACRA2, Butteraugli, and DSSIM are being developed to measure image quality algorithmically, as subjective testing is time-consuming and costly. These metrics can give an indication of image quality faster and cheaper than human testing.
• Quality Points: Different quality points are defined for lossy compression, ranging from visually lossless (SSIMULACRA2 = 90) to medium quality (SSIMULACRA2 = 60). The relevant range for the web in 2024 is medium to high quality, while for non-web use cases, it is high to (visually) lossless quality.
• Encoder Comparison: Different encoders like libjpeg-turbo, sjpeg, mozjpeg, jpegli, libavif, libjxl, and libwebp are tested for various qualities. JPEG XL is often the clear winner, compressing images better and faster than other codecs, especially for high quality to visually lossless images. At lower quality points, different encoders perform differently, and some like mozjpeg and jpegli offer better trade-offs.
• Decode Speed: Sequential JPEG is the fastest in terms of decode speed, followed by progressive JPEG. JPEG XL is somewhere in between. The decode speed of AVIF depends on the encoding method.
• Test Set Differences: The results vary depending on the test set. For web-sized images (about 1 megapixel each), the Pareto front is different from that for larger images. JPEG XL is the clear winner for larger images at high quality points.
• Overall Improvement: The new version of libjxl brings a substantial reduction in memory consumption and an improvement in speed, especially for multi-threaded lossless encoding. It consolidates JPEG XL's position as the best image codec currently available.