Adaptive Fused Prior Transfer for Controllable Generative Image Compression
AFP-GIC uses frozen AdaCode priors for controllable image compression with lower latency than DC-VIC.
AFP-GIC is a controllable generative image codec that transfers an adaptive fused prior from a frozen pretrained AdaCode model. Encoder-side fused-prior features shape the latent, while the decoder predicts a compatible prior from the compressed representation and control variables, so the prior is not transmitted. Versus DC-VIC, it reports 18.1% lower decoder latency and 31.10 million fewer inference parameters, a 20.5% reduction. On Kodak, CLIC2020, and DIV2K it remains competitive on PSNR and SSIM, with the clearest gains in NIQE and very-low-bitrate visuals.
- Decoder reconstructs a fused prior from the bitstream and control variables.
- Decoder latency falls 18.1% and inference parameters drop 31.10 million versus DC-VIC.
- Kodak, CLIC2020, and DIV2K results stay competitive on PSNR and SSIM.
- Largest perceptual gains show up in NIQE and very-low-bitrate images.
From the original article
1 min read at huggingface.co
Learned image compression achieves competitive rate-distortion performance, but very-low-bitrate reconstruction remains challenging because the transmitted representation cannot preserve fine textures and local structures. Perceptual and generative codecs synthesize missing details using reconstruction priors, while controllable codecs allow one model to cover different bitrate and reconstruction preferences.
By Yifei Pei, Ying Liu, Nam Ling · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.