💻 computer science

Owner3D: Ownership-Guided Style Writing for Training-Free Localized 3D Stylization

Owner3D is a training-free framework for localized 3D stylization that leverages ownership-guided style writing and boundary dual slots to inject reference styles strictly into target regions within large reconstruction models, thereby significantly reducing style leakage and preserving non-target appearances without requiring additional training.

Original authors: Suchang Tao, Kaifeng Shi, Zhiyan Liu, Zhuoyuan Jiang, Yuqi Ouyang

Published 2026-08-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Suchang Tao, Kaifeng Shi, Zhiyan Liu, Zhuoyuan Jiang, Yuqi Ouyang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an artist with a magic paintbrush that can instantly turn a plain 3D statue into a masterpiece. In the world of computer science, this is called "3D stylization." For a long time, artists could only paint the whole statue at once, or they had to painstakingly rebuild the statue from scratch for every new style. But recently, a new type of "Large Reconstruction Model" (LRM) appeared. Think of these models as super-fast chefs who can look at a few photos of an object and instantly whip up a complete, edible 3D model in a single step. The problem? These chefs are a bit clumsy with details. If you ask them to paint just the hat on a statue red, they often accidentally paint the face and the hands red too, because the "recipe" they use to remember the object's look is shared across the whole statue. This paper tackles that messiness, asking: How can we tell the chef, "Paint the hat, but leave the face alone," without having to retrain the chef or build a whole new kitchen?

The authors of this paper, a team from Sichuan University, introduce a clever new tool called Owner3D. Their main finding is that you can achieve precise, localized 3D painting without any extra training or slow, step-by-step rebuilding. They discovered that by giving the model a simple "ownership map"—a digital label that says "this part belongs to the hat, that part belongs to the face"—they can guide the painting process to stay exactly where it's supposed to.

Here is how Owner3D works, using a few fun analogies:

1. The "Ownership-Guided Style Writing"
Imagine the model's internal memory as a giant whiteboard where it writes down the object's features. Usually, when you ask for a new style (like a sunset pattern), the model writes that pattern all over the whiteboard, covering everything. Owner3D changes the rules. It acts like a strict editor who hands the model a "Write Mask." This mask is like a stencil that says, "You can only write the sunset pattern on the 'hat' section of the whiteboard." The model still uses the same single whiteboard (a "triplane"), but the editor ensures the new style only gets written where it belongs. This prevents the "style leakage" where the sunset accidentally spills onto the statue's face.

2. The "Boundary Dual Slots"
Sometimes, the line between the hat and the face is blurry. In the model's memory, the pixels right on the edge might be shared by both the hat and the face. If you only have one memory slot for that edge, the model gets confused and mixes the styles. Owner3D solves this by adding "Boundary Dual Slots." Think of this as a tiny, special drawer right at the edge of the hat. Inside this drawer, the model keeps two separate notes: one note for the hat's new style and another note for the face's original look. When the model needs to paint that tricky edge, it can peek into the drawer and choose the right note, ensuring the hat stays red and the face stays skin-colored, even if they are touching.

3. The "Surface-First Texture Readout"
Finally, when the model is ready to show you the final 3D object, it has to decide which color to show for every tiny dot on the surface. Sometimes, the model isn't sure if a dot belongs to the hat or the face because it can't see it clearly from all angles. Owner3D uses a "Surface-First" strategy. It's like a detective who trusts the most direct evidence first. If the model can see the dot clearly from the original photos, it uses that "surface" evidence to decide the color. If that's missing, it falls back to 3D clues, and then to the whiteboard clues. This hierarchy ensures the model picks the most reliable color, avoiding fuzzy or wrong colors in the final image.

The Results
The team tested Owner3D on a benchmark using 49 real-world 3D objects (like chairs and lamps) and 16 different style references. They found that their method was a huge improvement over existing tools. Compared to a method called StyleSplat, Owner3D reduced "appearance leakage" (where the style bleeds into the wrong areas) by 86.4%. Compared to another method called LAENeRF, it reduced leakage by 89.9%.

Perhaps most impressively, Owner3D does all this in about 31.10 seconds per object-style pair. Other methods that try to do similar 3D editing often take minutes or even hours, or they require the model to be retrained for every new object. Owner3D, however, works instantly with the pre-trained model, making it a fast and efficient way to edit 3D assets.

In short, Owner3D proves that you don't need to rebuild the entire engine to fix a single part of the car. By simply guiding the model with ownership maps and smart memory slots, you can paint specific parts of a 3D object with high precision, keeping the rest of the object perfectly intact. It's a step toward making 3D editing as easy and fun as painting on a tablet, but for real, interactive 3D worlds.

Technical Summary of Owner3D: Ownership-Guided Style Writing for Training-Free Localized 3D Stylization

Problem Statement

Localized 3D stylization aims to modify the appearance of a specific semantic part of a 3D object (e.g., a chair seat or roof) while preserving the geometry and appearance of the remaining surfaces. While Large Reconstruction Models (LRMs) enable efficient, feed-forward 3D asset generation from sparse images, applying localized stylization within this framework presents three critical challenges:

Style Leakage: In LRMs, style is injected into intermediate appearance representations before rendering. Consequently, final-image masking cannot prevent undesired style propagation to non-target regions.
Boundary Ambiguity: Compact triplane representations share appearance features across different surface regions. Near semantic boundaries, target and non-target parts compete for limited representation capacity, leading to ambiguous feature selection.
Incomplete Visibility: Determining the "ownership" of texture queries is difficult when visibility is incomplete, leading to unreliable feature selection during texture readout.

Existing methods, such as those injecting style keys into late cross-attention, are inherently global and struggle with precise part-level control. Other approaches relying on optimized scene representations or iterative optimization lack the efficiency and scalability of feed-forward reconstruction.

Methodology

Owner3D is a training-free framework that integrates localized appearance control directly into the LRM reconstruction process. It operates on a frozen LRM backbone (specifically InstantMesh-Large) and modifies the inference pipeline through three core modules:

1. Ownership-Guided Style Writing
This module restricts style injection to decoder queries associated with the target part.

Ownership Estimation: A content-only LRM pass reconstructs geometry to estimate surface ownership. Using SAM 3 for target segmentation and depth-consistency checks, a surface ownership score is calculated for each 3D point. This is projected onto triplane cells to create a "write mask" ( $W_k$ ) and a 3D ownership field ( $M_{3D}$ ).
Gated Style Injection: In the final six appearance-decoder layers, the framework computes a content-relative style residual ( $\Delta = Y^s - Y^c$ ), isolating the appearance change introduced by the style reference. This residual is gated by the ownership map ( $g_i$ ) before being added to the hidden state:
$H_i^{\ell+1} = \mathcal{F}_\ell(H_i^\ell + Y_i^{c,\ell} + \alpha g_i \Delta_i^{s,\ell})$
This ensures that style is written only to target-owned queries, producing a single localized stylized triplane without maintaining separate global style and content representations.

2. Boundary Dual Slots
To resolve the triplane capacity problem where target and non-target surfaces share the same local triplane support near boundaries, Owner3D introduces a sparse local capacity expansion.

Detection: Boundary locations are identified where the local neighborhood of a triplane cell contains both target-owned and preserve-owned regions.
Dual Storage: At these active boundary locations, the system maintains two separate feature sources: a "style slot" storing the locally stylized feature and a "preserve slot" storing the original reconstruction feature. This prevents the shared triplane cell from forcing a compromise between conflicting features.

3. Surface-First Texture Readout
During rendering, the framework hierarchically selects the most reliable appearance features for each surface query based on ownership evidence.

Priority Logic: The system prioritizes surface ownership (derived from visible source views) as the primary evidence. If unavailable, it falls back to the 3D ownership field ( $M_{3D}$ ), and finally to triplane write masks.
Feature Selection: Based on the determined ownership score and whether the query falls on a boundary slot, the renderer selects the appropriate feature: the stylized feature, the preserved feature, or the default feature (to avoid hard switching in ambiguous zones).

Key Contributions

The paper claims four primary contributions:

Owner3D Framework: A training-free localized 3D stylization framework enabling controllable part-level editing within a feed-forward reconstruction pipeline, significantly reducing non-target leakage compared to existing baselines.
Ownership-Guided Style Writing: A paradigm for LRMs that localizes reference-style injection while maintaining a single localized stylized triplane, avoiding the need for separate global style and content representations.
Boundary Dual Slots: A mechanism to overcome appearance ambiguity caused by shared triplane support, enabling independent local feature representation for target and non-target regions near boundaries.
Surface-First Texture Readout: A robust feature recovery strategy that prioritizes surface ownership evidence and progressively leverages 3D and triplane cues under incomplete visibility.

Experimental Results

The authors evaluated Owner3D on a benchmark constructed from Google Scanned Objects (GSO) and PartNet, involving 49 3D objects and 16 style references. The method was compared against Gaussian-splatting-based, NeRF-based, and image-space stylization methods.

Performance: Owner3D achieved the best Style Fidelity score (0.1545) and the lowest Leakage score (0.0042).
Leakage Reduction: Compared to StyleSplat and LAENeRF, Owner3D reduced appearance leakage by 86.4% and 89.9%, respectively.
Efficiency: The method is highly efficient, requiring only 31.10 seconds per object-style pair, outperforming other 3D baselines which often require minutes.
Ablation: Removing components like gated mixing or boundary dual slots resulted in significant degradation in style localization and increased non-target feature drift, validating the necessity of each module.

Significance and Claims

The paper positions Owner3D as a practical solution for controllable appearance editing in large reconstruction models. Its significance lies in:

Training-Free Operation: It achieves high-quality localized stylization without fine-tuning the LRM or training additional 3D models.
Integration of Ownership: It successfully integrates ownership guidance into the fixed reconstruction process to regulate pre-rendering style writing, a capability lacking in previous global injection methods.
Robustness: It addresses the specific challenges of triplane sharing and incomplete visibility, ensuring that non-target regions remain preserved even when target regions are stylized.

The authors conclude that Owner3D provides a superior balance of style fidelity, appearance preservation, and multi-view consistency, offering a scalable path for localized 3D editing. Future work is noted to explore improved region control and richer appearance representations.