Generative Image Compression

Generative image compression studies how to represent images with few transmitted bits while using a learned decoder to reconstruct missing detail. Approaches include entropy-coded latent and vector-quantized codes, semantic tokens, and diffusion or flow decoders, sometimes using reverse channel coding or shared randomness. Methods are compared on datasets such as Kodak, CLIC, and DIV2K using rate, pixel fidelity, perceptual metrics like LPIPS and DISTS, and distributional measures like FID. The central tension is that stronger generative priors can improve realism at extremely low bitrates but may alter source-specific details, while iterative decoding can be costly.