Overview of JPEG AI

JPEG AI (ISO/IEC 6048 | ITU-T T.840) is the first international image coding standard based on an end-to-end learning-based approach, developed jointly by ISO, IEC, and ITU-T under the Joint Photographic Experts Group (JPEG). Version~1 of the standard, targeting human visual consumption, was published as ITU-T T.840.1 / ISO/IEC 6048-1:2025 in 2025.

By creating the JPEG AI International Standard, the JPEG Committee has opened the door to more efficient and versatile image compression solutions that will benefit industries ranging from digital media and telecommunications to cloud storage and visual surveillance. This standard provides a framework for image compression in the face of rapidly growing visual data demands, enabling more efficient storage, faster transmission, and higher-quality visual experiences.

Scope

The scope of JPEG AI is the creation of a learning-based image coding standard offering a single-stream, compact compressed-domain representation, targeting:

• Human visualisation: significant compression efficiency improvement over image coding standards in common use at equivalent subjective quality;
• Machine consumption: effective performance for compressed-domain image processing and computer vision tasks, including image classification and semantic segmentation, without requiring full pixel decoding.

Target applications include cloud storage, visual surveillance, autonomous vehicles, image collection management, live visual monitoring, and media distribution. Moreover, the standard is designed to support a royalty-free licensing baseline.

Key Technical Features

• Superior compression efficiency: JPEG AI offers higher compression efficiency, leading to reduced storage requirements and faster transmission times compared to other state-of-the-art image coding solutions.
• Implementation-friendly encoding and decoding: JPEG AI codec supports a wide array of devices with different characteristics, including mobile platforms, through optimized encoding and decoding processes.
• Multi-branch decoding: three synthesis transforms (Main@Simple, Main@Base, Main@High) allow decoders to balance quality and complexity, enabling deployment across devices from smartphones to servers.
• Broad interoperability: designed for reproducible reconstruction across diverse hardware and software environments.
• Conditional colour separation: luminance and chrominance are coded independently, reducing peak memory usage and exploiting inter-channel correlation.
• Spatial random access: tile-based partitioning enables partial decoding of a coded picture.
• Progressive decoding: a lower-quality preview is obtainable from the hyperstream alone. The latent channels are ordered in order to improve the quality as more data is received.
• Wide-gamut and HDR support: multiple colour representations and rendering metadata (Dolby Vision, HDRVivid) are supported.
• Region-of-interest coding: spatially variable quantisation control via a 3D gain unit.

Proof-of-concept demonstration was shown on a Huawei Mate50 Pro, powered by a Qualcomm Snapdragon 8+ Gen 1, showcasing real-time decoding of high-resolution (4K) images. The demo highlighted key capabilities, including tiling support, full base operating point functionality, and decoding of images at arbitrary resolutions, thereby illustrating the practical feasibility of deploying JPEG AI on modern mobile devices.

Standard Parts

The first version of JPEG AI consists of five parts:

Performance

Building on mature deep learning technology, JPEG AI delivers significant rate reductions, approximately 30% for the same subjective quality relative to current best-performing conventional standards and introduces a compressed-domain representation that is inherently suited to machine-vision workflows without additional decoding overhead. Objective evaluation was done with multiple quality metrics, including MS-SSIM, FSIM, VIF, VMAF, PSNR-HVS, IW-SSIM, and NLPD. Subjective evaluation with the Double Stimulus Continuous Quality Scale (DSCQS) methodology confirmed that JPEG AI produces visually superior reconstructions from low to high bit rate ranges for natural photographic content.

History

The JPEG AI project was established at the 82nd JPEG meeting in January 2019. A Call for Proposals was issued in January 2022 (94th JPEG meeting); 14 registrations were received, with 12 submitted codecs evaluated for the standard reconstruction task. Subjective and objective evaluation was performed at the 96th JPEG meeting (July 2022). The collaborative standardisation phase culminated in the publication of Part 1 in 2025.