Flexible luma-chroma bit allocation in learned image compression for high-fidelity sharper images

High-fidelity learned image/video compression solutions are typically optimized with respect to l1 or l2 loss in RGB 444 format and evaluated by RGB PSNR. It is well-known that optimization of a fidelity criterion results in blurry images, which is typically alleviated by adding a content-based and/or adversarial loss terms. However, such conditional generative models result in loss of fidelity. In this paper, we propose a simple solution to obtain sharper images without losing fidelity based on learned flexible-rate coding using gained variational auto-encoder (gained-VAE) in the luma-chroma (YCrCb 444) domain. This allows us to implement image-adaptive luma-chroma bit allocation during inference, i.e., to increase Y PSNR at the expense of slightly lower chroma PSNR to obtain sharper images without introducing color artifacts based on the observation that Y PSNR correlates with image sharpness better than RGB PSNR. We note that the proposed inference-time image-adaptive luma-chroma bit allocation strategy can be incorporated into any VAE-based image compression model. Experimental results show that sharper images with better VMAF and Y PSNR can be obtained by optimizing models for YCrCb MSE with the proposed image-adaptive luma-chroma bit/quality allocation compared to stateof-the-art models optimizing RGB MSE at the same bpp.