Text-Driven 3D Stylization

• Text-driven 3D stylization is a method that transforms 3D assets using natural language prompts to alter visual styles and geometric details.
• It employs pretrained vision-language models (e.g., CLIP, diffusion models) and multi-view score distillation to ensure semantic alignment and consistency.
• Recent approaches focus on fine-grained part control, dynamic scene adjustments, and rapid feed-forward optimization for real-time applications.

Text-driven 3D stylization is the class of computational methods that generate or edit three-dimensional content (meshes, point clouds, vector sketches, radiance fields, or Gaussian splats) according to a user-supplied natural language prompt describing a desired visual style. The field encompasses workflows for static object stylization, articulated mesh animations, scene-level texturing, fine-grained part control, and sketch-based abstraction, all unified by the translation of textual descriptions into specific 3D visual, material, or structural outcomes. Modern systems achieve this by distilling priors from pretrained vision-LLMs (CLIP, diffusion models, GLIP), using gradient-based optimization or feed-forward neural architectures, and enforcing consistency and controllability across arbitrary viewpoints or motion sequences.

1. Foundations of Text-to-3D Stylization

Text-driven 3D stylization is fundamentally a cross-modal alignment problem: given a 3D representation $M$ and a natural language description $T$, the goal is to produce a stylized asset $M^*$ such that renderings from arbitrary views exhibit semantics and perceptual attributes matching $T$. The prevailing paradigm leverages pretrained image-text models (notably CLIP) and 2D or video diffusion models as frozen priors, supervising the stylization process through losses that maximize alignment between rendered images and target text or reference style images.

Key strategies include:

• Multi-view score distillation: Mapping text (and optionally style images) into embeddings; using diffusion models to generate denoising gradients; optimizing 3D parameters so that rendered images at diverse views minimize the misalignment in the embedding space (Chen et al., 29 Oct 2025, Kompanowski et al., 5 Jun 2024).
• Structural abstraction: For interpretable or lightweight outputs (e.g., 3D sketches or glyphs), sparse curve primitives or part-controlled Gaussians are employed, with optimizers acting on parameterized curves or splat clouds (Chen et al., 29 Oct 2025, Gan et al., 29 Nov 2025).
• Motion and 4D stylization: For dynamic content, text prompts may specify animation verbs (e.g., "flap", "move"), and supervision exploits pretrained video diffusion priors for temporal coherence (Chen et al., 29 Oct 2025, Youwang et al., 2022).

Crucially, zero-shot and training-free approaches dominate, with all semantic signal injected from frozen vision-LLMs and no need for curated paired 3D-text datasets.

2. Parametric Representations and Stylization Targets

A central technical challenge in text-driven 3D stylization is selecting a parametric representation that is both computationally tractable and expressive for the desired output domain. Major representation classes include:

Representation	Key Use Cases	Stylization Mechanism
Explicit polygonal meshes	Objects, scenes, animated bodies	Per-vertex MLPs, normal offsets
Neural radiance fields (NeRFs)	Volumetric/vista-level stylization	MLPs, tri-planes, color+density fields
3D Gaussian splats	Fast, scalable, dynamic scenes	Grouped latent editing, retraining
Bézier curves/sketches	Sparse, abstracted line drawings	Differentiable curve fitting, motion
Vectorized 3D strokes	Painterly/artist-style renderings	SDF-based, patch-level CLIP losses
Neural texture fields	UV-based scene/urban stylization	Hash-grid MLP, CLIP/Gram, class masking

• Explicit meshes: Most approaches (Text2Mesh, TANGO, X-Mesh, TeMO, 3DStyleGLIP) modify surface color and normal/geometry via MLP-based neural style fields, trained to maximize text-image similarity in rendered views (Michel et al., 2021, Chen et al., 2022, Ma et al., 2023, Zhang et al., 2023, Chung et al., 3 Apr 2024).
• NeRF and Gaussian splats: Methods for radiance fields or point clouds jointly optimize scene structure and appearance via hybrid contrastive and directionality objectives (NeRF-Art, CLIP3Dstyler), or use latent-diffusion editing in grouped splat space (GaussianBlender, Morpheus) (Wang et al., 2022, Gao et al., 2023, Ocal et al., 3 Dec 2025, Wynn et al., 3 Mar 2025).
• Vector sketches and strokes: Sparse, interpretable line-based representations are handled through differentiable Bézier or spline curves, with direct SDS-based gradient fitting for both geometry and animation trajectories (Chen et al., 29 Oct 2025, Duan et al., 2023).
• Semantic/part-aware and region-based control: For multi-object or urban scenes, segmentation (manual, geometric, or learned) enables per-part or per-region style control, either via part-specific prompts, loss mapping (TeMO, 3DStyleGLIP, SplatFont3D), or explicit region-aware losses (Improved 3D Scene Stylization) (Zhang et al., 2023, Chung et al., 3 Apr 2024, Gan et al., 29 Nov 2025, Fujiwara et al., 4 Sep 2025).

3. Loss Functions, Priors, and Optimization Protocols

Text-driven 3D stylization systems distill high-level semantic and stylistic intent into 3D geometry and appearance via several classes of losses:

• Semantic alignment losses: Typically,

$$\mathcal{L}_{\mathrm{CLIP}} = 1 - \cos ( E_{\text{text}}(T),\, E_{\text{img}}(I))$$

where $E_{\text{text}}$, $E_{\text{img}}$ are frozen encoders (Chen et al., 2022, Michel et al., 2021).

• Score Distillation Sampling (SDS): Denoising gradients from diffusion models applied to the output images as a function of the text prompt,

$$\mathbb{E}_{t, \epsilon} \big\| \epsilon_{\phi}(z_t; y, t) - \epsilon \big\|_2^2$$

possibly color-weighted, class-masked, or temporally structured (4D) (Chen et al., 29 Oct 2025, Kompanowski et al., 5 Jun 2024).

• Structure and geometric regularization: Losses on per-vertex displacements, stroke directionality, or compositional consistency; e.g.,

$$\mathcal{L}_{\rm geom} = \frac{1}{N} \sum_{i=1}^N \sum_{j=1}^2 \left\| \frac{\mathbf{p}_{i,j+1} - \mathbf{p}_{i,j}}{\|\cdot\|} - \frac{\mathbf{p}_{i,j} - \mathbf{p}_{i,j-1}}{\|\cdot\|} \right\|_2^2$$

(Chen et al., 29 Oct 2025).

• Style transfer losses: Mixes of Gram matrices, VGG-based perceptual losses, or sliced Wasserstein distances, optionally region- or class-masked for local effects (Chen et al., 16 Apr 2024, Fujiwara et al., 4 Sep 2025).
• Part/region-level style and GLIP-based alignment: For fine-grained control, localization and per-part matching of mesh/image regions and textual description are enforced using joint vision-language embeddings (Chung et al., 3 Apr 2024, Gan et al., 29 Nov 2025).
• Motion/Multi-frame and temporal losses: Temporal smoothness and shape preservation for animated or time-varying stylization (Chen et al., 29 Oct 2025, Youwang et al., 2022).
• Contrastive and directional losses: Relative and patch-based global-local contrastive alignments for robust style and geometry transformation (NeRF-Art, CLIP3Dstyler, X-Mesh) (Wang et al., 2022, Gao et al., 2023, Ma et al., 2023).

Optimization is predominantly gradient-based and operates directly on scene parameters, with training-free or single-asset tuning protocols that depend entirely on frozen backbone priors, making the approach dataset-agnostic.

4. Fine-Grained, Part-Aware, and Temporal Stylization

A major direction has been transferring control from holistic style application to per-part, per-object, or temporal stylization.

• Part-level and semantic stylization: 3DStyleGLIP uses GLIP's region-word alignment for identifying parts in rendered views, allowing distinct style sub-prompts for semantic components and loss enforcement at the part–prompt correspondence level (Chung et al., 3 Apr 2024). SplatFont3D applies dynamic component assignment and per-stroke text-prompted SDS, enabling independent stylization of font regions (Gan et al., 29 Nov 2025). TeMO achieves object-aware stylization by bipartite graph attention linking mesh point clusters and noun phrases, with multi-grained contrastive supervision for global and object-local fidelity (Zhang et al., 2023).
• Region-based style transfer in scenes: Scene-level methods (Text2Scene, Improved 3D Scene Stylization, StyleCity) leverage geometric or semantic segmentation and region-aware losses (multi-region IW-SWD, class-masked Gram, CLIP, or VGG) to deliver contextually consistent appearance changes to buildings, objects, or urban infrastructure (Hwang et al., 2023, Fujiwara et al., 4 Sep 2025, Chen et al., 16 Apr 2024).
• Time/coherence in animation: Frameworks such as 4-Doodle and CLIP-Actor leverage pretrained video diffusion model priors or pose/text-retrieval-based animation assembly, injecting temporal smoothness and shape-consistent motion for animated stylization (Chen et al., 29 Oct 2025, Youwang et al., 2022).

5. Scalability, Performance, and Practical Implementations

Scalability and inference efficiency have been challenging for text-driven 3D stylization due to the view-sampling cost, per-asset optimization time, and the need for high-frequency consistency. Recent advances address these issues:

• Feed-forward latent editors: GaussianBlender encodes grouped 3D Gaussians into geometry and appearance latents, editing only the appearance latent via a text-conditioned diffusion model for near-instant (<0.3s) stylization, preserving geometry and multi-view consistency (Ocal et al., 3 Dec 2025).
• Autoregressive/retargeting workflows: Morpheus stylizes temporally sampled RGBD frames with dual-masked diffusion then retrains 3D splats, enabling explicit control over shape and appearance strength (Wynn et al., 3 Mar 2025).
• Multi-stage and region-aware optimization: StyleCity and Improved 3D Scene Stylization decouple viewwise image stylization (with attention-sharing multi-view diffusion) from 3D structure retraining to ensure high-fidelity, regionally consistent results at city and scene scales (Chen et al., 16 Apr 2024, Fujiwara et al., 4 Sep 2025).
• User studies and quantitative evaluation: Cross-method benchmarking uses CLIP similarity, LPIPS, structure preservation, inference runtime, and rater-based scores for naturalness, style alignment, and detail diversity (see tables in (Ocal et al., 3 Dec 2025, Chen et al., 29 Oct 2025, Ma et al., 2023, Zhang et al., 2023, Fujiwara et al., 4 Sep 2025)).

6. Limitations, Open Problems, and Future Directions

• Semantic ambiguity and dataset bias: Reliance on CLIP or diffusion models induces ambiguities for abstract, composite, or underrepresented textual prompts (Chen et al., 2022, Chung et al., 3 Apr 2024). No existing method perfectly handles all real-world semantic part decompositions.
• Geometry–appearance disentanglement: While appearance-only style transfer is robust, concurrent geometry transformation (e.g., "making a mesh 'pixar-style'") is less interpretable and may induce unwanted deformations, with regularization balancing required (Wang et al., 2022, Wynn et al., 3 Mar 2025, Lee et al., 15 Aug 2025).
• Scene-scale or dynamic stylization: Consistent multi-object, regionally varying (e.g., foreground/background), and temporally evolving stylizations (4D) remain open challenges, handled only partially by recent methods (TeMO, StyleCity, Morpheus, CLIP-Actor) (Zhang et al., 2023, Chen et al., 16 Apr 2024, Wynn et al., 3 Mar 2025, Youwang et al., 2022).
• Evaluation and standardization: Objective, automated cross-dataset and multi-style benchmarking is emerging (MIT-30, Objaverse-3DStyle, GPTEval3D, user preference studies) but remains limited in scope and coverage (Ma et al., 2023, Yang et al., 2023, Kompanowski et al., 5 Jun 2024).
• Efficiency and interactivity: While GaussianBlender and Morpheus move toward real-time workflows, heavy per-instance optimization remains a bottleneck in many approaches (notably in mesh-based or volumetric representations) (Ocal et al., 3 Dec 2025, Wynn et al., 3 Mar 2025).

This suggests ongoing research will focus on more general, modular representations (latents, Gaussians, dynamic neural fields), more precise part and region-level linguistic conditioning, and further integration with efficient, off-the-shelf vision-language backbones for both generation and evaluation.

---

• (Chen et al., 29 Oct 2025) 4-Doodle: Text to 3D Sketches that Move!
• (Kompanowski et al., 5 Jun 2024) Dream-in-Style: Text-to-3D Generation Using Stylized Score Distillation
• (Chen et al., 2022) TANGO: Text-driven Photorealistic and Robust 3D Stylization via Lighting Decomposition
• (Youwang et al., 2022) CLIP-Actor: Text-Driven Recommendation and Stylization for Animating Human Meshes
• (Duan et al., 2023) Neural 3D Strokes: Creating Stylized 3D Scenes with Vectorized 3D Strokes
• (Zhang et al., 2023) TeMO: Towards Text-Driven 3D Stylization for Multi-Object Meshes
• (Chung et al., 3 Apr 2024) 3DStyleGLIP: Part-Tailored Text-Guided 3D Neural Stylization
• (Yang et al., 2023) 3DStyle-Diffusion: Pursuing Fine-grained Text-driven 3D Stylization with 2D Diffusion Models
• (Chen et al., 16 Apr 2024) StyleCity: Large-Scale 3D Urban Scenes Stylization
• (Ocal et al., 3 Dec 2025) GaussianBlender: Instant Stylization of 3D Gaussians with Disentangled Latent Spaces
• (Wynn et al., 3 Mar 2025) Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization
• (Gan et al., 29 Nov 2025) SplatFont3D: Structure-Aware Text-to-3D Artistic Font Generation with Part-Level Style Control
• (Michel et al., 2021) Text2Mesh: Text-Driven Neural Stylization for Meshes
• (Wang et al., 2022) NeRF-Art: Text-Driven Neural Radiance Fields Stylization
• (Ma et al., 2023) X-Mesh: Towards Fast and Accurate Text-driven 3D Stylization via Dynamic Textual Guidance
• (Hwang et al., 2023) Text2Scene: Text-driven Indoor Scene Stylization with Part-aware Details
• (Gao et al., 2023) CLIP3Dstyler: Language Guided 3D Arbitrary Neural Style Transfer
• (Fujiwara et al., 4 Sep 2025) Improved 3D Scene Stylization via Text-Guided Generative Image Editing with Region-Based Control
• (Lee et al., 15 Aug 2025) StyleMM: Stylized 3D Morphable Face Model via Text-Driven Aligned Image Translation

---

7. Schematic Overview of Key Approaches

Method	Representation	Primary Stylization Signal	Part/Animation Control	Notable Features
4-Doodle (Chen et al., 29 Oct 2025)	Bézier sketches	Multi-view SDS, video SDS	Explicit, via motion prompt	Training-free animation, structural constraint
Dream-in-Style (Kompanowski et al., 5 Jun 2024)	NeRF	Score distillation (style mixture)	Style image (reference)	On-the-fly style injection in diffusion
TANGO (Chen et al., 2022)	Mesh, analytic BRDF	CLIP loss, SG-based differentiable	No	SVBRDF+SG shading, fast photoreal stylization
X-Mesh (Ma et al., 2023)	Mesh	Dynamic text attention, CLIP loss	No	Fast convergence, automatic benchmarks
GaussianBlender (Ocal et al., 3 Dec 2025)	3D Gaussians	Feed-forward latent diffusion	Geometry/appearance disentangled	Instant inference, large-scale capability
3DStyleGLIP (Chung et al., 3 Apr 2024)	Mesh	GLIP alignment loss, CLIP alt.	Per-part style text	Joint part-localization and stylization
TeMO (Zhang et al., 2023)	Mesh, multi-object	Decoupled graph attention, contrast	Clustered by prompt nouns	Cross- and fine-grained semantic alignment
Morpheus (Wynn et al., 3 Mar 2025)	3D Gaussians	RGBD diffusion, Warp ControlNet	Appearance/geometry	Independent control, consistent retraining
SplatFont3D (Gan et al., 29 Nov 2025)	3D Gaussians	Part-weighted SDS, dynamic assign.	Stroke-level text prompts	Structure-aware font stylization
StyleCity (Chen et al., 16 Apr 2024)	Neural texture field	Semantic CLIP/Gram, class-masked L	Semantic mask, region	Progressive views, panoramic sky via diff.

---

In summary, state-of-the-art text-driven 3D stylization unifies pretrained vision-language priors, differentiable 3D representations, and multi-view/part-aware optimization or editing to produce geometry- and appearance-modified 3D content with strong semantic fidelity to natural language prompts.