Learned image compression methods using a generative decoder can reconstruct images at significantly higher perceptual quality than new hand-crafted codecs or other learned methods. Recently, diffusion models have been integrated into the decoding process to further enhance image quality.
However, the diffusion process is sensitive to several hyper-parameters, such as the number of steps, which are typically hard-coded and expected to perform well across various images. When applied to a single image, these parameters are often suboptimal.
In this work, we propose enhancing reconstruction quality by optimizing the diffusion process's decoding parameters for each image individually during encoding. This approach improves final quality with virtually no increase in bits-per-pixel. Additionally, we compare methods to minimize the additional computational impact during encoding.
We validate our approach on the CDC (Yang et al., 2024) and PerCo (Careil et al., 2023) image compression models using datasets like Kodak and DIV2K. Our results show clear improvements in LPIPS and PSNR without negatively impacting bits-per-pixel. This concept of optimizing quality tradeoffs can be readily applied to other diffusion-based image compression methods without the necessity of additional network training.