Facial editing is an important task with applications in entertainment, virtual reality, and digital avatars. Most existing approaches rely on generative models in the 2D image domain, while in 3D the task is typically performed through labor-intensive manual editing. We propose FaceParts, a framework for unsupervised segmentation and editing of Gaussian Splatting avatars. Unlike existing 2D or mesh-assisted methods, our approach operates directly in the Gaussian domain, decomposing avatars into semantically coherent facial parts without supervision. The method integrates feature disentanglement, density-based clustering, and FLAME-anchored part transfer, enabling precise editing and cross-avatar part swapping. Experiments on the NeRSemble dataset with 11 subjects demonstrate robust isolation of features such as beards, eyebrows, eyes and mustaches. Quantitative evaluation confirms that transferred segments adapt to pose and expression, while maintaining identity consistency (ID = 0.943), low Average Expression Distance (AED = 0.021) and low Average Pose Distance (APD = 0.004).
Figures & tables
Figure 1: Overview of the FaceParts reconstruction pipeline. The unsupervised segmentation process has three stages: Disentanglement , where a neural network groups similar Gaussians into shared facial components; Refinement , where density-based clustering extracts coherent subparts; and Extraction , where these parts are stored for reuse. The framework supports facial part swapping , enabling modification of a target avatar by transferring features (e.g., beard, eyes).
Figure 2: Framework overview: (a) Head avatars employ a hybrid solution of 3D Gaussians rigged to a parametric mesh model. (b) Reconstructing Gaussian features based solely on position encourages grouping of similar semantic regions within the same bottleneck channel. (c) Density-based clustering ensures that the components constitute smaller logical facial parts, such as an eyebrow. (d) The underlying FLAME model enables transferring facial parts without manual adjustments.
Figure 3: Examples of face-part transfer across identities. The top row shows extracted segments (beard, eyebrows, eyes, and mustache) from avatar 218, while the bottom row shows the same parts attached to avatar 210.
Figure 4: Dynamic behavior of face swapping. Top row: avatar 253 rendered across different poses and expressions. Bottom row: the same avatar with a beard transferred from avatar 218. The transferred segment adapts consistently to head pose and expression, following the deformations of the underlying FLAME mesh.
Figure 5: Segmentation results for different bottleneck sizes. Smaller bottlenecks (e.g., size 3 and 4) lead to interpretable groupings such as clothing, skin, and other features, while larger bottlenecks often produce fragmented clusters without clear semantic meaning.
Figure 6: Example highlighting the limitations of segmentation and generative models. (a) Segmentation, left: ground-truth image. Middle: output from a 2D face segmentation model. Right: our result. (b) Editing, middle: inpainting with a generative diffusion model. Right: our result. The segmentation model struggles to capture realistic appearance and fine-grained details, while the diffusion model lacks reliability in out-of-distribution cases (e.g., woman with a beard). Our approach preserves identity and structure while producing natural and consistent facial features.
Segment
AED ↓
APD ↓
ID ↑
FID ↓
Beard
0.037 (0.015)
0.007 (0.006)
0.905 (0.011)
–
Eyebrows
0.014 (0.006)
0.004 (0.003)
0.927 (0.007)
–
Eyes
0.004 (0.002)
0.001 (0.001)
0.986 (0.003)
–
Mustache
0.029 (0.014)
0.005 (0.004)
0.952 (0.010)
–
Global Avg
0.021 (0.009)
0.004 (0.003)
0.943 (0.008)
5.872
Table 1: Quantitative evaluation of face-part transfer. Metrics are averaged across all avatars and grouped by segment type. Results indicate that attaching an additional segment has only a small influence on expression and pose extraction. The largest variations are observed for the beard, likely because it is the largest segment. The overall FID score is also reported to quantify realism relative to unmodified avatars.
Figure 7: Comparison of transferring hair from avatar 306 to avatar 218 using our FaceParts and MEGA ( Wang et al., 2024a ) .
Figure 8: a) Lusage+Lsparsity b) Lsparsity c) Lusage d) no additional loss e) direct prediction on xyz coordinates f) prediction on xyz with the standard softmax. The sparsity loss encourages the network to cluster nearby regions, while the usage loss promotes a more uniform bottleneck distribution across the entire avatar. Models trained without HashGrid fail to disentangle smaller facial parts.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: The first row illustrates the overlap strategy with different values of parameter o . Higher values make the attached beard less visible, as the avatar’s Gaussians dominate. The second row shows the replacement strategy, where smaller values of parameter n result in a more visible beard.
Figure 10: Examples of facial part swaps.
Figure 11: First row: hair extracted from avatar 306 using the disentanglement network followed by clustering. Second row: face swap with an additional step. Beyond transferring Gaussians in relative coordinates, we locally adjust the FLAME mesh for avatar 218 (only the triangles corresponding to the hair region of avatar 306) for better alignment.
We present Edit3DGS, a unified framework for dynamic 3D head editing that integrates 2D instruction-guided diffusion with 3D Gaussian splatting. Unlike prior approaches that separately address frame-based edits or static 3D reconstruction, our method couples semantic controllability in the image domain with photorealistic, temporally consistent 3D representations. Given an input video, editable facial regions are masked and modified using a text-conditioned diffusion model to support fine-grained operations such as expression transformation, attribute modification, and appearance refinement. The edited frames are then aggregated through 3D Gaussian splatting to produce a coherent, high-fidelity avatar that preserves both identity and motion dynamics. To enforce consistency, Edit3DGS incorporates multi-view batch editing and lightweight inpainting strategies that recover lost expressions across timesteps. Experimental results demonstrate that our framework enables controllable, artifact-free head editing with smooth temporal transitions, offering practical applications in virtual avatars, immersive communication, film production, and interactive media.
Duy-Dat Tran, Trung-Nghia Le
University of Science, VNU-HCM, Ho Chi Minh, Vietnam · Vietnam National University, Ho Chi Minh, Vietnam
High-fidelity 3D head avatars have reached near-photorealistic quality. While recent methods enable text-driven manipulation, they struggle to provide fine-grained localized control, often entangling features or lacking geometric consistency. Modifying geometry through natural language currently requires slow per-prompt optimization or compromises identity and rigging. We present ManifoldSplat, the first end-toend framework for language-guided semantic shape editing of animatable 3D Gaussian Splatting avatars reconstructed from monocular videos. By performing edits within the structured FLAME manifold rather than directly optimizing an unstructured Gaussian cloud, we strictly preserve identity and animation. We introduce DeltaRegion, a per-region disentangled Conditional Variational Autoencoder (CVAE) delivering feedforward shape deltas, alongside a refining stage to recover view-consistent details. ManifoldSplat reconstructs and edits an avatar in ~90 seconds on a consumer GPU, rendering at ~800 FPS. Extensive evaluations demonstrate our approach sets a new state-of-the-art in localized prompt alignment, geometric coherence, and identity preservation. Project page and code: https://a-canela.github.io/manifoldsplat/
Antonio Canela, Jordi Sànchez-Riera
Universitat Politècnica de Catalunya · Institut de Robòtica i Informàtica Industrial, CSIC-UPC
We present ChromaGS, a method for real-time, language-guided color editing of animatable 3D Gaussian head avatars. Given a trained animatable avatar, users can instantly modify the color of semantic regions through natural language, with edits applied at render time and no retraining required. Our key insight is to augment each Gaussian primitive with learned soft assignments to semantic regions and decompose colors into region-level base colors and Gaussian-level residuals. This decomposition enables coherent color transfer: modifying a region's base color propagates naturally through all associated Gaussians while preserving fine appearance details encoded in residuals. A two-stage language pipeline translates text instructions into target colors, supporting both absolute specifications and relative adjustments. Unlike generative editing methods that may introduce unintended modifications, our approach provides deterministic, precisely localized semantic control. Experiments demonstrate faithful appearance preservation and intuitive interaction across diverse subjects. Project page and code are available at: https://a-canela.github.io/chromags/
Antonio Canela, Jordi Sànchez-Riera
Universitat Polit`ecnica de Catalunya · Institut de Rob`otica i Inform`atica Industrial, CSIC-UPC