LAGA: Layered 3D Avatar Generation and Customization via Gaussian Splatting

TL;DR

This work tackles the challenge of generating decomposable 3D clothed avatars by decoupling garments from the body and enabling garment-level edits. It introduces LAGA, a layered avatar framework based on Gaussian Splatting, featuring a coarse-to-fine garment generation strategy and a dual-SDS loss to enforce coherence between garments and other avatar components. To support garment transfer, the authors add three regularization losses—Human Fitting, Similarity, and Visibility—that guide Gaussian point movement across different body shapes. Extensive experiments show LAGA outperforms prior methods in texture and geometry quality, while enabling flexible garment customization and cross-avatar transfer, with promising implications for personalized avatars in gaming and VR. Overall, LAGA leverages the explicit, editable nature of Gaussian Splatting to achieve highly controllable, high-fidelity decomposable avatars.

Abstract

Creating and customizing a 3D clothed avatar from textual descriptions is a critical and challenging task. Traditional methods often treat the human body and clothing as inseparable, limiting users' ability to freely mix and match garments. In response to this limitation, we present LAyered Gaussian Avatar (LAGA), a carefully designed framework enabling the creation of high-fidelity decomposable avatars with diverse garments. By decoupling garments from avatar, our framework empowers users to conviniently edit avatars at the garment level. Our approach begins by modeling the avatar using a set of Gaussian points organized in a layered structure, where each layer corresponds to a specific garment or the human body itself. To generate high-quality garments for each layer, we introduce a coarse-to-fine strategy for diverse garment generation and a novel dual-SDS loss function to maintain coherence between the generated garments and avatar components, including the human body and other garments. Moreover, we introduce three regularization losses to guide the movement of Gaussians for garment transfer, allowing garments to be freely transferred to various avatars. Extensive experimentation demonstrates that our approach surpasses existing methods in the generation of 3D clothed humans.

Figures (7)

• Figure 1: We present LAGA, a novel layered avatar generation framework based on Gaussian Splatting (GS). With the layered structure, our generated clothed avatar can be decomposed to a human body with multiple individual garments, allowing users to assemble and edit specific garments to create new variations.
• Figure 2: Overview of the avatar component generation process in each layer. As outlined in the green box, our generation process of each layer mainly consists of three steps: (a) sparse initialization of Gaussian points, (b) density guidance to obtain coarse garment, (c) densification to obtain fine garment, In the beginning, based on the given layer's text description, we initialize a set of sparse Gaussian points using the parametric human model (SMPL) and associated joints. Then, these points are refined to approximate the broad shape of the target component in the coarse stage. Subsequently, in the fine stage, we densify the Gaussians to capture finer details and sharper features of the avatar component, aiming for high-quality results. To ensure coherence with other generated avatar components, a dual-SDS loss (as presented in the blue box) is introduced to optimize the Gaussian points in both coarse and fine stages. This loss function optimizes Gaussians from both local and global perspectives, enhancing the quality and coherence of the generated avatar component.
• Figure 3: Qualitative results. We compare our method with SOTA 3D human generators on six different prompts, each showing three camera views.
• Figure 4: Individual components for each avatar.
• Figure 5: Ablation for C2F. (a) Avatar w/ C2F. (B) Avatar w/o C2F.