Title: AvatarMix: Enabling Identity-Preserving Cross-Avatar Composition for Outfit Personalization

Abstract:

Current methodologies for transferring 3D avatar outfits encounter significant hurdles. Techniques that elevate two-dimensional edits into three dimensions frequently compromise the quality of both the outfit and the user's identity. Conversely, methods that treat body and clothing layers independently often result in visible intersection artifacts. To address these limitations, we present AvatarMix, a novel compositional framework that constructs avatars by directly combining the heads and bodies of two high-fidelity Gaussian avatars. This approach naturally maintains high outfit quality and eliminates intersection issues; however, it presents new difficulties in achieving a seamless integration point and preserving visual fidelity following body reshaping.

To overcome these obstacles, we developed a two-stage refinement process. First, SeamFix, a localized diffusion module, enhances the hair and neck regions to guarantee a flawless transition. Second, FullbodyFix serves as an optional full-body refinement step that restores garment appearance when retargeting causes degradation to the clothed body. Since both modules operate on renders from an already 3D-consistent Gaussian avatar, they significantly reduce multi-view artifacts compared to traditional 2D-to-3D lifting techniques.

Furthermore, to maintain the user's body identity, our mesh-based Gaussian representation allows for the integration of a robust mesh retargeting technique. This enables the precise reshaping of the clothed body to match the user's physique while effectively handling a wide variety of body shapes. Comprehensive experiments confirm that our method delivers state-of-the-art performance in terms of outfit fidelity and identity preservation, offering a fresh perspective on realistic 3D outfit personalization.

Project page: https://larsph.github.io/avatarmix/

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC