JPEG XL

• JPEG is the Joint Photographic Experts Group, which is the committee that designed the format.
• X is part of the name of several JPEG standards since 2000: JPEG XT, JPEG XR, JPEG XS.
• L means Long-term because the authors' intention for the format is to replace the legacy JPEG and last as long too.[3]

In 2017, JTC1/SC29/WG1 (JPEG) issued a Call for proposals for JPEG XL – the next generation image coding standard.[4]

The file format (bitstream) was frozen on December 25, 2020, meaning that the format is now guaranteed to be decodable by future releases.[5]

The main features are:[6][7]

• Improved functionality and efficiency compared to traditional image formats (e.g. JPEG, GIF and PNG);
• Image dimensions of over a billion (2³⁰-1) pixels on each side;[8]
• Up to 4100 channels i.e grayscale or RGB, optional alpha, and up to 4096 "extra" channels;[8]
• Progressive decoding (by resolution and precision);
• Lossless JPEG transcoding with ~20% size reduction;
  • CMYK JPEGs are not supported for transcoding, but they are very rare;[9]
  • Transcoding to progressive JPEG XLs is supported by the format but not yet implemented in the reference software;[10]
• Lossless encoding and lossless alpha encoding;
• Support for both photographic and synthetic imagery;
• Graceful quality degradation across a large range of bitrates;
• Perceptually optimized reference encoder;
• Support for wide color gamut and HDR;
• Support for animated content,
• Efficient encoding and decoding without requiring specialized hardware
  • In particular, JPEG XL is about as fast to encode and decode as old JPEG using libjpeg-turbo and an order of magnitude faster to encode and decode compared to HEIC with x265.[8] It is also parallelizable.
• Royalty-free format with an open-source reference implementation.[11]

JPEG XL codec architecture diagram

JPEG XL is based on ideas from Google's PIK format and Cloudinary's FUIF format (which was in turn based on FLIF).[12]

The format has a variety of encoding modes. On the legacy side, it has a mode that transcodes legacy JPEG in a more compact way for storage. On the more modern side, it has a lossy mode called VarDCT (variable-blocksize DCT) and a lossless/near-lossless/responsive mode called Modular which optionally uses a modified Haar transform (called "squeeze") and which is also used to encode the DC (1:8 scale) image in VarDCT mode as well as various auxiliary images such as adaptive quantization fields or additional channels like alpha. Both modes can use separate modeling of specific image features: splines, repeating "patches" like text or dots, and noise synthesis. Lossy modes typically use the XYB color space derived from LMS.[13]

Prediction is run using a pixel-by-pixel decorrelator without side information, including a parametrized self-correcting weighted ensemble of predictors. Context modeling includes specialized static models and powerful meta-adaptive models that take local error into account, with a signaled tree structure and predictor selection per context. Entropy coding is LZ77-enabled and can use both asymmetric numeral systems (ANS) and Huffman coding (for low complexity encoders or for reducing overhead of short streams).

It defaults to a visually near-lossless setting that still provides good compression.[8]

Animated (multi-frame) images do not perform advanced inter-frame prediction, though some rudimentary inter-frame coding tools are available:

• frames can update only parts of the canvas;
• besides replacing parts of the canvas, frames can also be blended, added or multiplied to parts of it;[14]
• up to four frames[15] can be remembered and referenced using the "patches" coding tool in later frames.[16]

• Official website
• Reference implementation on GitHub
  • builds:
    • Nightly dev builds
    • Windows Bin Builds (the page also contains many other programs)
• Community website