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

Antonio Canela; Jordi Sànchez-Riera

cs.CVOct 2, 2026

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

Authors: Antonio Canela, Jordi Sànchez-Riera

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

Abstract

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/

Figures & tables

Fig. 1: ChromaGS enables real-time, language-guided color editing of 3D Gaussian head avatars. Given a trained avatar (left), users can instantly modify semantic region colors through natural language instructions, with edits applied directly at rendering time, bypassing the need for per-edit optimization.

Fig. 2: (a) SAM3 extracts semantic masks from input frames. (b) Text-guided editing: Flan-T5-Small parses instructions into structured text; a resolver handles absolute colors (palette lookup via embedding similarity) and relative modifiers (HSV shifts). (c) Gaussians learn semantic logits (supervised by mask cross-entropy) and decompose color into base colors $\mathbf{b}_{k}$ plus residuals; editing replaces $\mathbf{b}_{k}$ while preserving fine details in residuals.

Fig. 3: Qualitative comparison. Results for skin, hair, and eyes edits across three subjects. Our method produces localized, natural color changes that preserve identity and fine details (e.g., specular highlights, skin texture). Both GaussianAvatar-Editor [ 5 ] and InstructPix2Pix [ 22 ] exhibit color bleeding into unrelated regions and over-saturated tones.

Method	Train $\downarrow$	Edit $\downarrow$	Render $\uparrow$
IPix2Pix [ 22 ]	N/A	$\sim$ 400 ms	N/A
GAEditor [ 5 ]	← $\sim$ 800 s →		$>$ 100 fps
ChromaGS (Ours)	$\sim$ 800 s	$\sim$ 50 ms	$>$ 100 fps

TABLE I: Runtime comparison. Our method requires a one-time

\sim

800 s training, after which any subsequent edit is applied instantly (

\sim

50 ms). In contrast, GaussianAvatar-Editor [ 5 ] relies on a coupled training-and-editing loop, requiring a full

\sim

800 s optimization for every new edit. InstructPix2Pix [ 22 ] requires no per-subject optimization as it processes each frame independently via diffusion.

Method	CLIP $\uparrow$	Identity $\uparrow$	Context $\uparrow$
IPix2Pix [ 22 ]	0.231	0.691	0.790
GAEditor [ 5 ]	0.229	0.863	0.790
ChromaGS (Ours)	0.212	0.986	0.980

TABLE II: Quantitative comparison with baselines. Our method achieves significantly better identity and context preservation. The baselines’ higher CLIP scores reflect a limitation of the metric itself: CLIP measures global image-text alignment and thus rewards visually dominant color shifts, even when they bleed into unintended regions.

Fig. 4: Ablation: semantic supervision. Without

\mathcal{L}_{\text{sem}}

, edits spread to unrelated areas: blue eyes tint the hair and tongue, blonde hair bleeds into the face and clothing, and red lipstick affects the skin color as well. Our full model produces clean, confined edits.

Method	CLIP $\uparrow$	Identity $\uparrow$	Context $\uparrow$
Ours w/o $\mathcal{L}_{\text{sem}}$	0.198	0.969	0.967
Ours	0.213	0.990	0.986

TABLE III: Quantitative ablation: Semantic supervision. Without

\mathcal{L}_{\text{sem}}

, color changes spread beyond the target region, degrading all three metrics. Notably, CLIP also drops: unlike the diffusion baselines that apply intense, globally visible color shifts, the unguided leakage here dilutes the edit rather than amplifying it.

3D Scene Editing

3D scene editing studies how to make controlled changes to objects, appearance, layout, and motion while keeping a scene coherent across views.

See all 88 papers

ManifoldSplat: Language-Guided Semantic Shape Editing of 3D Gaussian Head Avatars
Oct 2, 2026Antonio Canela, Jordi Sànchez-Riera3D Shape Editing 3D Gaussian Splatting
Universitat Politècnica de Catalunya · Institut de Robòtica i Informàtica 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
COSY: Compositional 3DGS Synthesis for Disentangled Human Head Editing
May 22, 2026Florian Barthel, Shalini De Mello, Koki Nagano +33D Gaussian Splatting Disentangled Representation Learning
Fraunhofer Heinrich-Hertz Institute, Germany · Humboldt University Berlin, Germany · NVIDIA, USA

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

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

Antonio Canela, Jordi Sànchez-Riera

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

3D Shape Editing 3D Gaussian Splatting 3D Avatar Generation

Open paper AskPDF arXiv

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/

Abstract

3D Scene Editing

Explore similar work

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

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

COSY: Compositional 3DGS Synthesis for Disentangled Human Head Editing