Learned Image Compression

Learned image compression trains neural encoders and decoders to represent images compactly, then uses quantization and learned entropy models to code those representations efficiently. Common approaches include convolutional or transformer-based codecs with hyperpriors and autoregressive context models, alongside methods that adapt bit allocation to image regions or downstream tasks. On benchmarks such as Kodak and CLIC, learned codecs are compared with standards like VVC using rate-distortion measures, perceptual metrics, and task accuracy; diffusion-based decoders also target realistic reconstructions at extremely low bitrates. A central tension is balancing source-faithful detail against perceptual realism or machine-task performance, while keeping encoding, decoding, and model costs practical.