Quality and performance benchmarking of learned image codecs with conventional image codecs

Abstract

Learned image codecs (LICs) have demonstrated strong rate-distortion performance in academic benchmarks, but their readiness for production deployment remains unclear. In this paper, we present a comprehensive benchmarking study comparing several state-of-the-art learned image codecs — including NNVC, JPEG AI, and the DCVC family — against established conventional codecs such as JPEG, AV1, AV2, and VVC, using production-oriented encoding configurations. Evaluation is conducted on approximately 300 images from five public datasets spanning diverse resolutions and content types. We assess compression efficiency using BDRATE computed against JPEGLI across 11 quality metrics, including both conventional measures (PSNR, SSIM, VMAF, SSIMULACRA2) and perceptual metrics (LPIPS, DISTS). We also report encoding runtime and computational complexity (kMACs/pixel) for GPU-based learned codecs. Our results show that DCVC-RT achieves the highest overall compression efficiency (~52% BDRATE gain over JPEGLI) while encoding at ~30 ms per image on GPU — approximately 3,000× faster than VTM at comparable quality. However, the gain over the best conventional codecs (AV2, VVC) narrows to ~3%.