AFP-GIC slashes decoder latency by 18.1% and cuts 20.5% of parameters for low-bitrate image compression

Learned image codecs have made great strides, but very-low-bitrate reconstruction still suffers from lost textures and local details. Perceptual and generative codecs address this with learned reconstruction priors, and controllable variants let a single model cover multiple bitrate and quality preferences. However, controllability alone doesn't solve the decoder-side prior problem: under severe bit constraints, the decoder must infer missing details from limited transmitted information, and existing codebook-based designs rely on single-codebook token-based priors that are often insufficient.<br><br>Enter AFP-GIC (Adaptive Fused Prior Transfer for Controllable Generative Image Compression). The method leverages a frozen pretrained AdaCode model to create an adaptive fused prior. At the encoder, fused-prior features guide the latent representation; at the decoder, a lightweight network predicts a compatible fused prior from the compressed representation and selected control variables, enabling prior-guided reconstruction without transmitting the prior itself. The authors show the fused-prior family contains single-codebook choices as special cases and provide a reconstruction-error upper bound analysis. Under a unified benchmark, AFP-GIC reduces decoder latency by 18.1% and overall parameter count by 31.1 million (20.5%) relative to DC-VIC. Experiments on Kodak, CLIC2020, and DIV2K demonstrate competitive PSNR, with the clearest perceptual gains in NIQE scores and visual comparisons at very low bitrates. The paper is submitted to IEEE Access and code is available on arXiv.