Representation Analysis for 3D Avatars

Abstract

Recent generative models have enabled the creation of photorealistic 3D avatars from a limited number of images. However, the impact of generation settings on representation quality and computational cost has not been sufficiently studied. In this work, instead of proposing a new avatar generation method, we analyze the effect of different generation settings in a CAP4D-based 3D avatar pipeline. The results show that medium-cost configurations, where the generation time is reduced to approximately one quarter, can preserve visual quality close to high-cost settings. PSNR, LPIPS, CSIM and JOD results indicate that these configurations should be interpreted not as a strict quality improvement, but as an efficient cost–quality trade-off for practical use.