Adaptive Fused Prior Transfer for Controllable Generative Image Compression
The paper’s key move is to transfer a fused reconstruction prior from a frozen AdaCode model without sending that prior in the bitstream.
AFP-GIC uses encoder-side prior features to shape the latent representation, while the decoder predicts a compatible fused prior from the compressed data and control variables. The authors argue that better prior alignment tightens a reconstruction-error bound, and that their fused-prior setup contains single-codebook designs as special cases. One pretrained model is evaluated across five bitrate points. Against DC-VIC, it reports 18.1% lower decoder latency and 31.10 million fewer inference parameters, with strongest gains showing up in NIQE and very-low-bitrate visual comparisons. HF Daily Papers' note
AFP-GIC uses encoder-side prior features to shape the latent representation, while the decoder predicts a compatible fused prior from the compressed data and control variables. The authors argue that better prior alignment tightens a reconstruction-error bound, and that their fused-prior setup contains single-codebook designs as special cases. One pretrained model is evaluated across five bitrate points. Against DC-VIC, it reports 18.1% lower decoder latency and 31.10 million fewer inference parameters, with strongest gains showing up in NIQE and very-low-bitrate visual comparisons. HF Daily Papers' note
score 4