We introduce Gaussian Wardrobe, a novel framework to digitalize compositional 3D neural avatars from multi-view videos. Existing methods for 3D neural avatars typically treat the human body and clothing as an inseparable entity. However, this paradigm fails to capture the dynamics of complex free-form garments and limits the reuse of clothing across different individuals. To overcome these problems, we develop a novel, compositional 3D Gaussian representation to build avatars from multiple layers of free-form garments. The core of our method is decomposing neural avatars into bodies and layers of shape-agnostic neural garments. To achieve this, our framework learns to disentangle each garment layer from multi-view videos and canonicalizes it into a shape-independent space. In experiments, our method models photorealistic avatars with high-fidelity dynamics, achieving new state-of-the-art performance on novel pose synthesis benchmarks. In addition, we demonstrate that the learned compositional garments contribute to a versatile digital wardrobe, enabling a practical virtual try-on application where clothing can be freely transferred to new subjects. Project page: https://ait.ethz.ch/gaussianwardrobe
Figures & tables
Figure 2: Gaussian Wardrobe digitalizes compositional neural avatars from multi-view videos. Our pipeline consists of two major components: (left) a compositional Gaussian representation and (right) a framework for learning neural garments. We first reconstruct a mesh template M from the first video frame and segment it into body Mb , garment templates M{u,ℓ} in the zero-shaped canonical space. During training, each layer learns a separate U-Net F to predict the parameters of 3D Gaussian primitives M from pose-conditioned positional maps P . We composite the 3D Gaussians G from all layers and render RGB images I^ and segmentation masks S^ to compute the training loss L . The learned neural garments are shape-agnostic and can seamlessly transfer to other subjects for avatar virtual try-on.
Figure 3: Exemplar of avatar virtual try-on. Given a reconstructed avatar (left), we replace its lower garment with a new skirt, Mℓ′ . The combined avatar can be animated to a novel pose θ′ , which may introduce minor penetration artifacts (middle). We resolve these artifacts on-the-fly during rendering with our online penetration detection algorithm (right).
Method
PSNR ↑
SSIM ↑
LPIPS ↓
4D-DRESS [ 62 ]
Animatable Gaussians [ 32 ]
27.75
0.9577
0.0531
LayGA [ 33 ]
27.58
0.9574
0.0543
Gaussian Wardrobe (Ours)
28.06
0.9579
0.0527
ActorsHQ [ 23 ]
Animatable Gaussians [ 32 ]
27.92
0.9411
0.0418
Table 1: Evaluation of novel pose synthesis on 4D-DRESS and ActorsHQ datasets. Best results are highlighted in bold . Our method consistently outperforms all baselines on all evaluation metrics ( cf . Fig. 4 ).
Figure 4: Qualitative comparisons of novel pose synthesis on 4D-Dress and ActorsHQ datasets. Our method can better model the dynamics of free-form garments, such as skirts (top) and vests (middle), and generate realistic renderings with sharper facial and garment details. In contrast, the baseline methods suffer from artifacts, such as blurry faces and semi-transparent clothing, and fail to reproduce fine details like wrinkles or pockets.
Method
mIoU ↑
Recall ↑
F1-Score ↑
w/o Lsg
0.848
0.898
0.924
w/o Lreg
0.879
0.922
0.940
w/o Lpe
0.883
0.927
0.942
Ours Full
0.893
0.936
0.947
Table 2: Ablation study of the loss terms. We assess the effectiveness of each loss term from the full objective. We rendered segmentation masks of garment layers and compared them against the ground-truth masks ( cf . Fig. 5 ). The results show that the full model achieves the best performance.
Figure 5: Visualization on the effectiveness of loss terms. We rendered segmentation masks of garment layers to visualize the impact of each regularization term. The evaluation shows that our full model produces cleanest results, while the absence of Lpe , Lreg and Lsg leads to self-penetration or irregular segmentation.
Figure 6: Importance of segmentation loss in virtual try-on. We visualize the results of virtual try-on using the variants our method trained with and without Lsg . The garment’s appearance becomes corrupted due to the entanglement of body and clothing.
Figure 7: Virtual try-on application. Our method enables flexible 3D virtual try-on across different subjects. The resulting avatars can be animated to a novel pose.
Figure 8: Visualization of penetration-aware rendering. Penetration-aware rendering removes small artifacts between garment layers.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Animatable Gaussians [ 32 ]
30.51
0.9491
0.0332
LayGA [ 33 ]
30.36
0.9482
0.0337
Gaussian Wardrobe (Ours)
30.86
0.9489
0.0331
Table 3: Quantitative evaluation of novel view synthesis on ActorsHQ dataset. We use a held-out camera view to evaluate the performance of novel view synthesis.
Method
4D-DRESS-00127-Inner
4D-DRESS-00185-Inner
4D-DRESS-00127-Outer
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
AG [ 32 ]
0.9579
28.73
0.0503
0.9612
28.78
0.0485
0.9539
25.62
0.0609
LayGA [ 33 ]
0.9579
28.63
0.0506
0.9607
28.53
0.0502
0.9533
25.49
0.0626
Ours
0.9585
28.98
0.0502
0.9615
29.15
0.0478
0.9535
25.95
0.0604
Method
ActorsHQ-Actor05
ActorsHQ-Actor08
ActorsHQ-Actor01
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
SSIM ↑
PSNR ↑
LPIPS ↓
Table 4: Full quantitative evaluation of novel pose synthesis on the 4D-DRESS and ActorsHQ dataset.
Figure 9: Importance of diffused skinning fields. To deform the template mesh into the canonical pose, we evaluate different strategies for querying skinning weights in inverse LBS. Nearest-neighbor querying fails to preserve topology near the armpits, whereas our diffused skinning fields produce smoother and more accurate garment meshes.
Figure 10: More results on virtual try-on. Our method supports flexible 3D try-on across different subjects, and the generated avatars can be animated into novel poses.
Figure 11: Results on ActorsHQ. Our method produces photorealistic 360° rendering under novel views and poses.
Figure 12: Results on ActorsHQ. Our method produces photorealistic 360° rendering under novel views and poses.
Reconstructing animatable 3D human avatars from a single image remains particularly challenging for loose garments, whose geometry and motion cannot be adequately represented by body-aligned topology and skinning. We present Forwardrobe, a feed-forward framework for reconstructing garment-aware Gaussian avatars from a single image. Forwardrobe explicitly separates clothing from the body in canonical Gaussian space and equips the garment layer with continuity-aware geometry and skinning initialization, pose-conditioned non-rigid deformation, and appearance adaptation. These designs improve garment reconstruction and visual quality during animation, particularly for skirts and dresses. The separated garment layer additionally forms an independently controllable 3D asset, enabling garment editing, transfer, and 3D virtual try-on. Experiments demonstrate improved garment reconstruction quality and greater flexibility in garment manipulation compared with existing single-image avatar reconstruction methods.
Daisheng Jin, Shuyun Wang, Ying He
Nanyang Technological University · Singapore · The University of Queensland +1
Existing 3D clothed avatar reconstruction methods achieve high visual fidelity but ignore geometric structure and physical plausibility. They either model clothed humans as a single deformable surface or attempt garment disentanglement without enforcing geometric constraints, resulting in ambiguous garment boundaries and no control over stacking or layer ordering. To address these limitations, we introduce DAMA (Disentangled body-Anchored Gaussians for Controllable Multi-layered Avatars), a 3D avatar reconstruction method that produces physically plausible clothed avatars through a dedicated representation and reconstruction method. At the representation level, we bind Gaussians to SMPL-X faces using barycentric in-plane coordinates and a positive normal offset. Based on this parameterization, the reconstruction method lifts 2D segmentations to body-anchored Gaussians, refines layers using topology-guided correction, and jointly optimizes geometry and appearance. DAMA is the first Gaussian avatar reconstruction method from multi-view images to achieve physically plausible layering, clean garment separation, and explicit stacking control. On the full 4D-DRESS dataset (82 scans), it achieves state-of-the-art performance in geometry reconstruction, garment separation, penetration rate, and penetration depth. The representation further supports user-defined garment reordering and fast conversion of body-conforming garments to simulation-ready meshes. Project Page: https://danieleskandar.github.io/dama/
Daniel Eskandar, Berna Kabadayi, Garvita Tiwari +1
University of Tübingen, Germany · Tübingen AI Center, Germany · Zuse School ELIZA, Germany +2
The generation of high-fidelity, animatable 3D human avatars remains a core challenge in computer graphics and vision, with applications in VR, telepresence, and entertainment. Existing approaches based on implicit representations like NeRFs suffer from slow rendering and dynamic inconsistencies, while 3D Gaussian Splatting (3DGS) methods are typically limited to static head generation, lacking dynamic control. We bridge this gap by introducing AGORA, a novel framework that extends 3DGS within a generative adversarial network to produce animatable avatars. Our key contribution is a lightweight, FLAME-conditioned deformation branch that predicts per-Gaussian residuals, enabling identity-preserving, fine-grained expression control while allowing real-time inference. Identity is further preserved through spatial shape conditioning of the identity branch, and expression fidelity is enforced via a dual-discriminator training scheme leveraging synthetic renderings of the parametric mesh. AGORA generates avatars that are not only visually realistic but also precisely controllable. Quantitatively, we outperform state-of-the-art NeRF-based methods on expression accuracy while rendering at 250 FPS on a single GPU and, notably, at ∼9 FPS under CPU-only inference -- to our knowledge the first demonstration of CPU-only animatable 3DGS avatar synthesis. This work represents a significant step toward practical, high-performance digital humans. Project website: https://ramazan793.github.io/AGORA/
Ramazan Fazylov, Sergey Zagoruyko, Aleksandr Parkin +2
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Polynome AI, Dubai, UAE · MTS AI, Moscow, Russia