The core of practical learning-based image compression

For practical learning-based image compression, the key design choices that jointly optimize perceptual quality and runtime were systematically compared. Several new techniques were also included in ablation experiments, with a focus on finding combinations that are actually deployable.

The core was performance-aware neural architecture search. Millions of backbone configuration candidates were explored to select a model that maximizes compression performance based on perceptual metrics while satisfying on-device runtime constraints.

The resulting new codec significantly improved the balance between speed and perceptual quality. In rigorous subjective user evaluations, it achieved 2.3-3x bitrate reduction compared to AV1, AV2, VVC, ECM, and JPEG-AI, and an additional 20-40% reduction compared to the best existing learning-based codecs.

Speed is also practical. On an iPhone 17 Pro Max, encoding a 12MP image took 230ms and decoding took 150ms, which is faster than most top ML-based codecs running on a V100 GPU.