ManifoldSplat: Language-Guided Semantic Shape Editing of 3D Gaussian Head Avatars

Antonio Canela; Jordi Sànchez-Riera

cs.CVOct 2, 2026

ManifoldSplat: Language-Guided Semantic Shape Editing of 3D Gaussian Head Avatars

Authors: Antonio Canela, Jordi Sànchez-Riera

Organizations: Universitat Politècnica de Catalunya · Institut de Robòtica i Informàtica Industrial, CSIC-UPC

Abstract

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/

Figures & tables

Figure 2: Overview of ManifoldSplat. Top: the ManifoldSplat pipeline turns a monocular video and a prompt into an animatable edited avatar via a frozen DeltaRegion edit followed by Gaussian training and warped refining (Sec. 3.4 ). Bottom: DeltaRegion, our text-conditioned shape-editing CVAE; the region gate zeroes out non-selected latent blocks before decoding (Sec. 3.3 ).

Method	CLIP-D $\uparrow$	ArcFace $\uparrow$	$TV_{2D}$ $\downarrow$
InstructPix2Pix [ 2 ]	0.0432	0.5212	0.0210
GaussianAvatar-Editor [ 24 ]	0.0467	0.7600	0.0126
ManifoldSplat (Ours)	0.0490	0.9319	0.0122

Table 1: Full-avatar editing quantitative comparison.

Variant	CLIP-D $\uparrow$	ArcFace $\uparrow$	$TV_{2D}$ $\downarrow$
ManifoldSplat (full)	0.0490	0.9319	0.0122
w/o (a) Surface Consist.	0.0481	0.9278	0.0123
w/o (b) Smooth-Color tanh	0.0431	0.9318	0.0125
w/o (c) Warped Refining	0.0470	0.9335	0.0131
w/o (a)+(b)+(c)	0.0437	0.9322	0.0132

Table 2: Ablation of the 3DGS enhancements.

Method	Reconstruction $\downarrow$	Edit $\downarrow$	Render $\uparrow$
IPix2Pix [ 2 ]	N/A	$\sim$ 400 ms	N/A
GAEditor [ 24 ]	← $\sim$ 800 s →		$>$ 100 fps
ManifoldSplat (Ours)	$\sim$ 60 s	$\sim$ 30 s	$>$ 800 fps

Table 3: Runtime comparison.

Figure 3: Full-avatar editing comparison on NeRSemble [ 16 ] against InstructPix2Pix and GaussianAvatar-Editor.

Figure 4: Shape editing comparison . Meshes are color-coded by per-vertex displacement from the neutral identity; only DeltaRegion (Ours) keeps the displacement concentrated on the ears.

Figure 5: Fine-grained local edits. Illustrates ManifoldSplat’s ability to realize a wide variety of fine-grained shape changes within a single facial region.

Figure 6: Ablation of the 3DGS enhancements on “thicker lips” . Red boxes mark the characteristic failure modes that each component prevents (floaters and stray colored splats).

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Region	CLIP-D $\uparrow$	ArcFace $\uparrow$	$TV_{2D}$ $\downarrow$
Face	0.0440	0.8879	0.0121
Nose	0.0498	0.9329	0.0122
Eyes	0.0525	0.9060	0.0122
Ears	0.0745	0.9741	0.0123
Neck	0.0448	0.9642	0.0123
Lips	0.0282	0.9262	0.0123

Appendix

Table 4: Per-region quantitative results of ManifoldSplat.

Figure 7: Per-vertex displacement heatmaps for representative DeltaRegion edits. Color encodes the

\ell_{2}

norm of each vertex’s displacement between the edited and unedited FLAME meshes (blue

\to

red), normalized per mesh. Each subfigure is labeled with the prompt fragment fed to DeltaRegion. All heatmaps are computed using the source identity of subject 104_EXP_5 from the NeRSemble dataset [ 16 ] .

Figure 8: Extended qualitative results. This figure illustrates ManifoldSplat’s ability to realize a wide variety of fine-grained shape edits across diverse subjects and attributes.

Figure 9: Fine-grained nose edits under non-frontal views on an in-the-wild sequence. (a) the unedited reconstructed avatar and (b-d) three fine-grained nose edits, all rendered from (1-2) two near-profile viewpoints that best expose the nasal geometry.

3D Shape Editing

3D shape editing changes the geometry, appearance, or parts of digital objects using spatial controls, prompts, images, or generative models.

See all 53 papers

ChromaGS: Text-Driven Semantic Editing of 4D Gaussian Avatars
Oct 2, 2026Antonio Canela, Jordi Sànchez-Riera3D Scene Editing 3D Gaussian Splatting
Universitat Polit`ecnica de Catalunya · Institut de Rob`otica i Inform`atica Industrial, CSIC-UPC
Edit3DGS: Unified Framework for Dynamic Head Editing via 2D Instruction-Guided Diffusion and 3D Gaussian Splatting
Jun 16, 2026Duy-Dat Tran, Trung-Nghia Le3D Scene Editing Video Editing
University of Science, VNU-HCM, Ho Chi Minh, Vietnam · Vietnam National University, Ho Chi Minh, Vietnam
AGORA: Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars
Dec 6, 2025Ramazan Fazylov, Sergey Zagoruyko, Aleksandr Parkin +23D Gaussian Splatting 3D Asset Generation
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Polynome AI, Dubai, UAE · MTS AI, Moscow, Russia

Oct 2, 2026J/K move · Enter open · S save

ChromaGS: Text-Driven Semantic Editing of 4D Gaussian Avatars

Antonio Canela, Jordi Sànchez-Riera

Universitat Polit`ecnica de Catalunya · Institut de Rob`otica i Inform`atica Industrial, CSIC-UPC

3D Scene Editing 3D Gaussian Splatting

Open paper AskPDF arXiv

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/

Abstract

3D Shape Editing

Explore similar work

ChromaGS: Text-Driven Semantic Editing of 4D Gaussian Avatars

Edit3DGS: Unified Framework for Dynamic Head Editing via 2D Instruction-Guided Diffusion and 3D Gaussian Splatting

AGORA: Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars