Better image compression at ultra-low bitrates using adaptive priors

Very-low-bitrate image compression struggles to preserve fine textures and local structures because the decoder has too little transmitted information to reconstruct missing details. This work proposes AFP-GIC, which uses an adaptive fused prior—learned patterns from a frozen pretrained model—to guide reconstruction at both encoding and decoding stages without transmitting the prior itself. The decoder predicts a compatible fused prior from the compressed bitstream and control variables, enabling a single model to handle variable bitrate and quality preferences. On standard benchmarks (Kodak, CLIC2020, DIV2K), AFP-GIC achieves competitive PSNR and stronger perceptual quality (measured by NIQE), with 18% lower decoder latency and 20.5% fewer parameters than the baseline DC-VIC.