Source-linked AI summary
OrnaStyler: Ornament-Aware Latent Editing for Content-Preserving 3D Stylization
Tomohiro Aizawa, Shigeru Kuriyama, Chunzhi Gu
TL;DR
Fine-grained ornament-aware 3D stylization must add style-specific structure without losing an asset’s geometry and identity. OrnaStyler performs staged, inversion-guided editing of geometry and appearance, and reports state-of-the-art performance in content preservation, style fidelity, and visual realism.
Problem
Fine-grained ornamentation remains difficult because stylization must preserve source geometry and object identity while integrating new style-specific details.
Method
OrnaStyler is a training-free framework that uses flow inversion for ornament-enhanced voxel synthesis and adjacency-aware feature inpainting for appearance harmonization.
Results
OrnaStyler achieves state-of-the-art ornament-aware 3D editing across content preservation, style fidelity, realism, and overall visual plausibility.
Takeaways & Limitations
Explicitly modeling ornament spatial configuration supports stylization that balances added structure with content consistency and coherent appearance.
Takeaways & Limitations
Because OrnaStyler relies on TRELLIS, detailed designs such as text may be blurred or poorly preserved when the backbone struggles with high-frequency textures.
Abstract
from arXiv · showhide
Text-guided style editing of 3D assets is essential for adapting existing objects to diverse visual aesthetics in digital content creation. Despite rapid progress in 3D shape modeling, faithfully stylizing an existing asset remains challenging when the desired stylization involves fine-grained structural ornamentation, which requires the model to preserve the source geometry and object identity, while coherently integrating new style-specific details. We propose \textbf{OrnaStyler}, a zero-shot framework for text-guided ornament-aware 3D stylization. Built upon rectified flow-based generative modeling, OrnaStyler introduces an inversion-guided editing strategy that recovers content-aware latent representations at both geometry and appearance levels in a staged manner to facilitate faithful editing. Our core idea is to explicitly model the spatial configuration of stylistic elements, thereby mitigating the fundamental tension between content preservation and style expression in the voxel space. Specifically, at the geometry level, we manipulate voxel representations through flow inversion to synthesize ornament-enhanced structures while preserving the spatial identity of the source asset. Then, at the appearance level, we introduce an adjacency-aware feature inpainting mechanism to harmonize newly generated ornaments with the original content, yielding coherent geometry-appearance integration. Our approach operates solely in the inference phase and enables selective editing over geometric augmentation or appearance stylization. Extensive experiments on both generated and real-world 3D assets against prior methods demonstrate that OrnaStyler achieves state-of-the-art editing performance in terms of content preservation, style fidelity, and overall visual realism. Code is available at: https://github.com/tomohiro0427/OrnaStyler
1. Introduction
OrnaStyler addresses the challenge of adding style-specific structural ornaments while preserving an asset’s geometry and identity. It uses inversion-guided, two-level editing to coordinate ornament geometry and appearance, achieving strong preservation, fidelity, and realism.
- 3D stylization must introduce structural ornamentation while preserving source geometry and object identity, a challenge that remains underexplored.
- Existing UV-texture and 2D-edit-to-3D-generation paradigms can produce inconsistent appearance, geometric artifacts, or deviations from the original content.
- OrnaStyler models ornamentation as spatial organization around existing content and uses a two-stage voxel- and appearance-level editing paradigm.
- Flow inversion recovers a content-aware seed latent, enabling style-conditioned ornament-enhanced voxels while adaptively localizing ornaments and limiting corruption of the content shape.
- Adjacency-aware 3D feature inpainting harmonizes ornament and content features, addressing distortions near their boundaries.
- The training-free framework achieves state-of-the-art editing effectiveness across realism, content consistency, and style fidelity.
2. Related work
Prior 3D stylization methods commonly edit appearance through 2D or texture-centric representations, while OrnaStyler uses a 3D structure latent-guided paradigm. The method overview separates ornament geometry synthesis from appearance stylization.
- Earlier 3D generation methods use point clouds, meshes, SDFs, diffusion models, GANs, or distilled 2D diffusion priors.
- Image-guided approaches optimize NeRF-based color and appearance to match a reference style while attempting to preserve original content.
- OrnaStyler’s overview presents a 3D structure latent-guided approach that first synthesizes ornament-enhanced voxel coordinates, then generates latent features for appearance stylization.
- Texture-centric methods are limited in modifying underlying geometry, and newer 3D-prior approaches still struggle with fine-grained color patterns and local designs.
3. Preliminaries
TRELLIS represents 3D assets with structured latents containing voxel geometry and appearance features. Its staged rectified-flow generators produce these components before an object decoder synthesizes the asset.
- TRELLIS represents geometry and appearance jointly with a structured latent Z={P,F}, where P contains voxel coordinates and F contains active-voxel features.
- Voxel coordinates index a K-resolution grid, while each active voxel has a d-dimensional latent feature.
- Rectified flow perturbs data toward Gaussian noise through linear interpolation and learns a time-dependent velocity field for the reverse process.
- The conditional flow-matching objective trains the velocity field to match the displacement from clean samples toward noise.
- TRELLIS first generates sparse structural information with G_S, then generates latent features with G_L conditioned on the voxel coordinates.
- The structured latent is passed through an object decoder to produce 3D assets in representations such as NeRF, 3D Gaussian Splatting, or meshes.
4. Methods
OrnaStyler edits 3D assets in stages, first generating ornament-aware geometry through inversion-guided, region-aware voxel synthesis, then harmonizing appearance with latent feature editing and adjacency-aware inpainting.
- OrnaStyler starts from the content asset and edited prompt, aiming to preserve the source geometry and appearance while adding style-specific ornamentation.
- Flow inversion-guided ornament stylization: Flow inversion recovers content-aware latent representations, while delayed inversion depth preserves the source structure during geometry editing.The method initializes inversion from encoded content voxels and uses a partially inverted latent because full inversion can erase fine-grained geometry.
- Flow inversion-guided ornament stylization: Region-aware structure stylization combines content- and edited-prompt velocity fields so occupied content regions remain anchored while unoccupied regions receive stylistic augmentation.The decoded result is merged with the original content voxels to resolve potential structural discontinuities or holes.
- Latent-level style editing: At the appearance level, OrnaStyler stylizes content features, re-noises them, and uses feature-level inpainting to harmonize ornament features with the edited content.Full trajectory inversion is used for appearance features because the geometric structure is already fixed.
- Latent-level style editing: Adjacency-aware masking identifies ornament voxels and neighboring content regions by dilating ornament coordinates and measuring their overlap, guiding seamless feature integration.The resulting mask specifies where local latent features should be synthesized while preserving the edited content feature.
5. Experiments
Experiments compare OrnaStyler with prior 3D style-editing approaches using qualitative, quantitative, perceptual, selective-editing, control, ablation, and inpainting analyses. The results support strong content preservation, coherent ornamentation, flexible editing control, and improved style–content integration.
- Evaluation: Qualitative comparisons evaluate OrnaStyler against 2D-edit-3D-generation and UV-texture-based style-editing methods.The evaluation edits rendered images before 3D reconstruction for the first category and uses UV-texture approaches for the second.
- Quantitative evaluation: OrnaStyler consistently outperforms compared methods across all content-preservation metrics.The evaluation renders content and edited assets from 12 fixed viewpoints and measures SSIM, PSNR, LPIPS, CLIP, and FID; the comparison primarily focuses on content preservation.
- User study: Participants consistently prefer OrnaStyler over competing methods across style fidelity, ornamentation quality, and overall naturalness.The study contrasts five methods and reports that OrnaStyler produces visually plausible and structurally faithful edits.
- Selective editing: Selective editing separates texture-only, ornamentation-only, and full editing, with each mode affecting distinct aspects of stylization.Texture-only editing changes appearance without geometric distortion, ornamentation-only editing adds spatially coherent geometry while preserving appearance, and combining both provides full editing.
- Control: OrnaStyler exposes spatial and semantic control over stylization through dilation-based intensity adjustment and semantic ornament attributes.Larger dilation numbers expand feature propagation and strengthen stylization, while different style descriptions can control ornament attributes.
- Analysis: Ablations show that region-aware guidance, content guidance, delayed injection, and adjacency-aware inpainting support natural, content-consistent ornamentation.Removing these components causes spatial incoherence, excessive ornamentation, uncontrolled propagation, or degradation of fine-grained geometry and texture details.
6. Limitation
OrnaStyler inherits limitations from the TRELLIS backbone and currently supports only additive ornament editing. Fine details such as text may blur, while subtractive geometric edits remain unsupported.
- Backbone limitations: TRELLIS’s representational capacity limits OrnaStyler’s fidelity for high-frequency details such as text.The paper attributes blurred or poorly preserved details to TRELLIS’s difficulty reconstructing high-frequency textures and suggests more expressive backbones as future work.
- Editing scope: OrnaStyler supports additive ornament editing but not subtractive operations that remove original content geometry.The paper gives erosion as an example of an unsupported subtractive operation.
7. Conclusion
OrnaStyler models ornament spatial configuration for zero-shot text-guided 3D editing. It uses flow inversion and adjacency-aware masking to preserve structure while harmonizing content and style.
- OrnaStyler is a zero-shot text-guided 3D asset editing framework that explicitly models ornament spatial configuration.
- Flow inversion traces seed latents at voxel and appearance levels to prevent structural corruption during stylization.
- An adjacency-aware spatial mask improves visual consistency at content–ornament boundaries.
- Extensive qualitative and quantitative experiments report state-of-the-art ornament-aware 3D shape-editing performance against prior methods.
A. Setting and evaluation metrics
Evaluation renders each generated sample from 12 fixed camera poses. Content similarity is assessed with SSIM, PSNR, and LPIPS.
- Setting: Each generated sample is rendered from 12 fixed camera poses along a circular trajectory at 30° elevation.Azimuths are uniformly spaced every 30° across [0°, 360°).
- Evaluation metrics: SSIM measures structural correspondence between original and edited rendered images using luminance, contrast, and structural information.
- Evaluation metrics: PSNR measures pixelwise fidelity through the ratio between maximum signal and reconstruction noise, while LPIPS measures deep-feature perceptual distance.
B. Efficiency analysis.
OrnaStyler adds computational overhead for ornament-aware processing but maintains a favorable runtime relative to the compared methods.
- OrnaStyler incurs overhead from voxel augmentation and feature harmonization because it builds on the TRELLIS framework.
- Despite this overhead, OrnaStyler achieves the second-best runtime among all compared approaches.
- The runtime result indicates a favorable trade-off between ornament-aware processing and computational efficiency.
C. Parameter sensitivity
The parameter study examines delayed injection across thin and dense structures using content consistency and ornament coverage. It also reports robustness to prompt imprecision and semantic conflict, alongside user-study procedures.
- Parameter sensitivity: The delayed injection study varies α within [0.5, 1.0] on helicopter, desk lamp, cart, and sofa assets representing thin and dense structures.
- Parameter sensitivity: Smaller α limits ornamentation, whereas larger α increases ornament coverage at the cost of content consistency; the method therefore uses α=0.8.
- Parameter sensitivity: The study evaluates the trade-off with CLIP similarity for content consistency and Surface Coverage Ratio for ornament coverage.
- Prompt types: OrnaStyler remains visually consistent with incomplete or inaccurate prompts and coherently applies snow styling to a fire-themed asset.
- User study: The user study includes 30 participants completing 20 questions with three evaluation criteria per question in sessions lasting approximately 20–25 minutes.
F. Asset attribution
This section lists Sketchfab assets used by OrnaStyler and provides the algorithms for voxel-level ornament synthesis and latent-level style editing.
- F. Asset attribution: Sketchfab assets are used under the Creative Commons Attribution 4.0 International license, with creator attributions listed for each model.Examples include “Asylum Bed,” “Tiny Tikes Kids Table,” and “Iron Rack Shelf for Home Storage.”
- F. Asset attribution: Algorithm 1 describes voxel-level ornament synthesis and outputs stylized coordinates from content and style inputs.The procedure uses denoising iterations, a delay rate, and guidance strength.
- F. Asset attribution: Algorithm 2 describes latent-level style editing and outputs a stylized asset from edited and style text inputs.It uses a pre-trained TRELLIS model and a specified number of denoising iterations.
- F. Asset attribution: The paper states that the overall procedures for both stages are provided in Algorithms 1 and 2.
H. More results
The paper provides additional qualitative editing examples and visual comparisons with prior methods on generated and real-world asset datasets.
- H. More results: Additional style-editing results of OrnaStyler are presented in Figure 15.The figure uses colored textual prompts for style editing and black texts for the generated content assets.
- H. More results: Figure 16 visualizes comparisons against prior methods on Trellis-Generated data.
- H. More results: Figure 17 visualizes comparisons against prior methods on Sketchfab data.