Apple Announces New Image Compression Technology “PICO” on GitHub

Apple Announces New Image Compression Technology “PICO” on GitHub

Apple has announced PICO (Perceptual Image Codec), a machine-learning-based image codec built from the ground up to be optimized for the human visual system. The project was published on GitHub via Apple’s machine learning research page (apple.github.io/ml-pico) and accompanied by a research paper on the preprint platform arXiv (arXiv:2605.05148), released around May 23–25, 2026. No specific timeline for practical deployment in Apple products has been announced.

A Codec Built Around Human Perception

Traditional image codecs have long measured quality by pixel-by-pixel similarity to the original — a metric that does not always reflect what humans actually perceive as beautiful or natural. PICO takes a fundamentally different approach: its model was developed based on large-scale subjective user research, training using a combination of pixel-matching loss, perceptual quality loss functions, GAN-based adversarial loss, and specialized loss functions designed to prevent degradation of fine details such as small text and tile boundaries.

The “GAN-based loss” component trains the codec to make reconstructed images appear more realistic rather than merely mathematically close to the original — a subtle but meaningful distinction for everyday photographs and graphics.

Performance: How Does PICO Stack Up?

Based on human-rated perceptual quality scores from Apple’s large-scale subjective studies, PICO claims substantial bitrate savings over every major modern codec:

Against the best existing machine-learning codec alternatives, PICO additionally claims a 20% to 40% further bitrate reduction — a significant margin even in a field that has seen rapid progress from neural compression research groups worldwide.

On-Device Speed: Faster Than a Server GPU

Perhaps the most striking claim is around processing speed. Apple benchmarked PICO on the iPhone 17 Pro Max, where it encodes a 12-megapixel image in 230 milliseconds and decodes it in 150 milliseconds. Apple notes this is faster than most top ML-based codecs running on a server-grade NVIDIA V100 GPU, making PICO genuinely practical for on-device use without relying on cloud infrastructure.

Cross-Platform Robustness

Unlike most learned codecs — which can produce inconsistent outputs across different hardware, operating systems, or compiler environments — PICO comes with explicit cross-platform robustness guarantees. This means that the same encoded image should decode identically regardless of the device or platform, a critical requirement for any codec intended for real-world deployment rather than research demonstration.

Apple achieved this through extensive architectural choices discovered across the search of millions of model configurations, jointly optimizing both perceptual quality and on-device runtime throughout the design process rather than as an afterthought.

What’s Next?

The research is publicly available on Apple’s GitHub and the arXiv paper, and the project page at apple.github.io/ml-pico includes an interactive comparison tool allowing side-by-side evaluation of PICO against other codecs. However, Apple has not announced any plans for integrating PICO into its operating systems, developer frameworks, or existing image formats such as HEIC. The codec currently handles lossy compression only, though this is consistent with the image quality use cases it targets.

Whether PICO becomes a next-generation iPhone photo format, a web delivery codec, or remains a research artifact for now remains to be seen — but as the first learned codec combining this level of perceptual benchmarks, on-device speed, and cross-platform deployment guarantees, it marks a notable milestone in practical image compression.