GOATex: Geometry & Occlusion-Aware Texturing

We thank the reviewer for recognizing the novelty of our work and for the constructive feedback on improving the clarity of our presentation.

---

Q1: The hit level and its relationship with residual face cluster and normal flip are the core of your paper, which should be demonstrated more clearly rather than being placed in the supplement. This makes it difficult to follow.

A1: We appreciate the feedback and will revise the main paper layout to more clearly present this pipeline in the core sections, rather than relying on the supplementary file.

---

Q2.The author has claimed to be the first work to simultaneously generate internal and external textures of objects, and the texture generation part relies on SyncMVD. Therefore, it is sufficient to compare the baseline SyncMVD and present more ablation studies. There is no need to compare with other texture generation models, as they completely collapse. Alternatively, such content can be moved to the supplement.

A2: Thank you for the constructive feedback. We agree that a more detailed comparison with SyncMVD, along with thorough ablation studies, is essential to highlight the core contributions of our occlusion-aware framework.

To address this, we conducted additional experiments during the rebuttal period by constructing a series of ablated models, where components were cumulatively added on top of SyncMVD. This setup allows both step-by-step ablation and direct comparison with SyncMVD.

For evaluation, we conducted A/B preference tests using SyncMVD as the reference, and each ablated variant as well as our full model as the comparison targets. To improve robustness, we employed multiple GPT-based evaluators (GPT-4o-mini, GPT-4o, GPT-4.1, and GPT-o3), and ran each configuration four times to account for response variability. In the table below, each score represents the win rate (%) of the corresponding method over SyncMVD.

While we observe overall performance gains across stages for the advanced GPT models, a slight drop appears when the residual face clustering is applied without normal flipping and backface culling. This is expected, as residual clustering alone cannot fully resolve occlusion. In some cases, two geometrically adjacent faces may be grouped into the same hit level, even if one fully occludes the other from all external views. In such cases, the occluded face cannot be textured during its designated rendering stage, and once that stage completes, it is excluded along with the rest of the level, resulting in missed interior content.

This limitation is mitigated in the final step by further applying normal flipping and backface culling, which introduces an indirect peeling-off effect. Instead of removing previously textured faces, we flip their normals and apply view-dependent culling—making them invisible from certain camera viewpoints. This increases the chance that previously occluded faces, which share the same hit level, are no longer blocked and can be exposed and textured in subsequent renderings. As a result, the method produces more complete and semantically coherent textures, especially in tightly occluded interior regions.

It is also worth noting that while our final textured renderings show clear improvements in the complete pipeline, intermediate ablated methods may not always exhibit visually striking differences in the final outputs. This is because early-stage outputs often produce out-of-distribution (OOD) inputs to the depth-conditioned ControlNet, leading to suboptimal or unstable generations. As a result, the benefits of each component are most clearly observed in intermediate representations, such as depth maps and UV maps, that reveal differences in geometric completeness and structural coherence at each hit level. Therefore, to provide a comprehensive and intuitive understanding of each stage, we will also incorporate side-by-side comparisons of depth maps, UV layouts, and textured renderings in the revised ablation analysis.

---

Q3.The ablation study uses complex car cases. Simple examples should also be presented to facilitate understanding of how different modules work.

A3: As discussed in A2, we will incorporate a range of examples, from simple to complex, in the revised manuscript to more clearly illustrate the role of each module in our framework. These examples will be supported by step-by-step visualizations using depth maps, UV maps, and textured renderings, making the effect of each component easier to interpret. We believe this will enhance the clarity of the ablation study and help readers better understand how different modules contribute to the final results.

---

Q4: Can a superface be assigned to different hit levels?

A4: No, a superface is strictly assigned to a single hit level.

Assigning a unique hit level to each superface allows our method to progressively reveal and texture occluded geometry in a controlled, layer-by-layer manner. In contrast, assigning multiple hit levels to a single superface would break this peeling-off progression, making it difficult to expose and texture occluded regions properly.

We sincerely thank the reviewer for the detailed and constructive feedback. We believe the combination of precise editorial suggestions and thoughtful technical insights will significantly contribute to improving the clarity, depth, and completeness of the work.

---

Q1: About the code release

A1: Absolutely! We are happy to release the full codebase upon publication to support reproducibility and encourage further research.

---

Q2: About alternative approaches that do not rely on pre-trained image models and use different representations.

A2: Thank you for the suggestion. We agree that mentioning such alternative methods would make our related work more comprehensive. We will revise the manuscript to include a brief discussion of these representations in the Related Work section.

---

Q3: Minor suggestions on clarifying notation in Fig. 2, repositioning Figures 3 and 4, correcting a method name in Fig. 5, and fixing a typo.

A3: We sincerely appreciate your attentive and thoughtful comments. We will reflect suggested edits in the revision.

---

Q4: The number of participants recruited in the user study.

A4: The number of participants is reported in the supplementary material, Section A.3.1, line 58. For clarity, we summarize them here:

• vs TEXTure: 21 participants
• vs SyncMVD: 33 participants
• vs Paint3D: 30 participants
• vs TEXGen: 45 participants

---

Q5: Conducting FID/KID-based evaluation by comparing against real or generated interior images.

A5: Thank you for the suggestion.

Following the reviewer’s recommendation, we conducted an FID-based evaluation using pseudo-reference images synthesized with depth-conditioned Stable Diffusion (SD 1.5), as the ground-truth meshes lack realistic interior textures. Notably, all evaluated methods are also based on SD 1.5.

To generate pseudo-references, we rendered depth maps from the ground-truth meshes (per view and hit level) and used depth-conditioned SD 1.5 to synthesize reference images. These were used to compute FID scores against the corresponding outputs of each method, separately for exterior (hit level 1) and interior regions (hit levels ≥ 2). We averaged scores across views and meshes:

While lower Interior FIDs of TEXTure and SyncMVD may appear favorable, they are misleading: (1) TEXTure leaves interiors blank or smooth, with occasional exterior texture “bleed” through mesh holes; SyncMVD extrapolates exterior colors via Voronoi heuristics. Neither produces semantically meaningful, prompt-aligned interiors. (2) The pseudo-references for hit levels ≥ 2, synthesized from single-view depth, lack multiview consistency and occlusion reasoning, often blending exterior and interior content, unlike MVD-based approaches in SyncMVD and ours.

These low Interior FIDs largely reflect (1) exterior-like texture continuity in methods without actual interior synthesis, and (2) pseudo-references biased toward such appearances. In contrast, methods that generate semantically meaningful interiors may deviate from these references and be unfairly penalized. We will include this analysis and qualitative examples in the revised manuscript.

Regarding exterior FID: Although our method scores higher, this is not a core limitation. GOATex specifically targets the challenging task of synthesizing textures for occluded interiors (hit-level ≥ 2), which prior work does not address. As GOATex is modular by design, high-fidelity exterior rendering (hit-level 1) can be integrated using existing methods. Our contribution thus complements, rather than replaces, current pipelines.

---

Q6: Concern about potential bias in the qualitative evaluation due to imbalance between interior and exterior images.

A6: Thank you for raising this important point.

To address the concern regarding evaluation balance, we revised both the user study and GPT-based evaluation setups to ensure fair and unbiased comparisons between interior and exterior texturing. Specifically, following the visualization style used on our project page, we rendered for each method:

• one GIF showing the exterior of the object, and
• one GIF showing the interior via a cut-away or sliced view.

These side-by-side visualizations were shown to participants or GPTs, providing one exterior and one interior view per method per object. Due to time constraints, we report results from 16 human raters who passed a vigilance test, with plans to expand this pool further. To compensate for this, we included four GPT variants, GPT-4o-mini, GPT-4o, GPT-4.1, and GPT-o3, and repeated each evaluation four times to reduce variance and improve robustness.

Below are the aggregated preference rates (%) for GOATex over each baseline:

While GOATex received strong and consistent preference from human raters across all comparisons, its relative advantage diminished when compared to TEXTure and SyncMVD in evaluations conducted with GPT-4 variants. We elaborate on this observation in A5.

---

Q7: About disagreement patterns between human preference and GPT, and cases where other methods perform better.

A7: Thank you for the interesting question.

To better understand human-GPT disagreement patterns, we conducted both quantitative and qualitative analyses:

1. Correlation with human ratings (Pearson’s r; r = 1 indicates perfect correlation):

• GPT-4o-mini: 0.22
• GPT-4o: 0.31
• GPT-4.1: 0.43
• GPT-o3: 0.34

2. Agreement metrics (Cohen’s κ, averaged over κ > 0; κ = 1 indicates perfect agreement):

• GPT-to-GPT: 0.54
• User-to-user: 0.31
• User-to-GPT: 0.27

These quantitative results indicate that:

• Advanced GPT models (especially GPT-4.1 and GPT-o3) show stronger alignment with human raters.
• GPT models exhibit higher internal consistency than human raters.
• Notably, in 17 cases, user-to-GPT agreement exceeded κ = 0.5, and in two cases, GPT-o3 achieved perfect agreement (κ = 1.0) with a human rater, demonstrating that GPTs can align closely with human preferences in specific contexts.

3. Qualitative analysis further revealed a consistent pattern of divergence:

• Humans weighed both interior and exterior quality and particularly penalized incomplete or semantically incoherent or incomplete interiors.
• GPTs tended to favor sharper exterior textures, even when interiors were missing or incorrect — a bias consistent with FID-based evaluation discussed in A5.

This explains why GPTs often underrated our method: GOATex is designed for occluded interiors, whose quality humans appreciate more than GPTs, which emphasize visible exterior realism. We will include representative examples and full results in the revised manuscript to better contextualize these differences.

---

Q8: About the computational cost of the full pipeline.

A8: Thank you for the insightful suggestion.

We systematically analyzed the runtime of our pipeline across meshes with increasing face counts by progressively subdividing mesh faces of five representative assets. For each mesh, we measured the number of faces and superfaces, time spent on superface construction, hit-level assignment, and MVD-based rendering & synthesis. The table below shows the average across the assets:

(*Hit-level assignment is reusable when generating multiple variants.)

As the reviewer correctly pointed out, the number of mesh faces has a significant impact on runtime. In our Objaverse 1.0-based dataset, however, mesh sizes are typically modest (median: 8.5k faces), keeping total runtime practical (under ~10 minutes in most cases).

We also find that the number of superfaces grows sublinearly with mesh size. While the pipeline is designed to operate at the superface level, most steps currently rely on face-level processing due to the reliance of existing implementations. Moving to superface-level processing throughout is a promising direction for reducing runtime on high-res meshes.

We will include this analysis and a corresponding runtime plot in the revision.

---

Q9: Request for a deeper discussion of the limitations of the proposed method beyond what's noted in the submitted version.

A9: Thank you for the insightful question.

One key limitation of GOATex lies in its semantically inconsistent hit-level assignment. Our current decomposition relies solely on geometric visibility computed via multi-view ray casting, which can result in semantically coherent regions being split across different hit levels (as also noted by Reviewer nix7 in Q4). While this does not lead to noticeable seams or texture artifacts in our experiments, it highlights a structural limitation of relying purely on geometry. Incorporating semantic-aware grouping into the hit-level assignment could enable more fine-grained texturing by allowing the pipeline to operate at the level of semantic entities. This would make it possible to assign different prompts to distinct parts, opening the door to richer, more controllable synthesis beyond the current progressive outside-in approach.

We will include this discussion in the revised manuscript to clarify the limitation and motivate future extensions.

We sincerely thank the reviewer for recognizing the importance and novelty of the proposed task. We appreciate the encouraging feedback on our occlusion-aware approach and the overall evaluation.

---

Q1: The approach struggles to effectively handle complex 3D meshes comprising multiple layers.

A1: We thank the reviewer for raising this point.

Scene complexity arises from various factors, including the number of semantic entities, their spatial arrangement, and the underlying geometry. As the number of semantic entities increases, the mesh tends to become more layered and entangled, making it increasingly difficult to disambiguate structures and generate proper textures across them. Furthermore, a high face count also complicates UV mapping, which can negatively affect texture quality and consistency.

In our current setup, the evaluated meshes span a broad range of semantic and geometric complexity, from simple to highly intricate cases. While it is not straightforward to define or quantify the number of "layers" in a scene, each mesh contains on average about 1.35 semantic entities, with some containing up to 7. The geometric complexity also varies significantly, with a median face count of 8.5K and a maximum of 166K.

Qualitatively, our method is demonstrated to perform well even on meshes with such multiple internal structures and layers. For instance, in the car example shown in Figure 6 of the main paper, components such as the steering wheel and seats are relatively small yet semantically important. While these parts are challenging to isolate and texture individually, our method (GOATex) is able to produce plausible textures, whereas baseline methods fail to generate meaningful results in such settings. Additionally, as illustrated in row 3, column 1; row 6, column 1; and row 6, column 3 of our project page, our approach generates coherent and prompt-aligned textures even in scenes with multiple internal or layered components. This is further supported by our quantitative evaluation, where our method is strongly preferred over baselines by both human raters and GPT-based assessments.

That being said, we acknowledge that even more compositionally rich and deeply nested configurations could be imagined. For example, consider a children's bedroom scene containing a closed toy chest filled with plush dolls, a bunk bed with built-in drawers and under-bed compartments, and a ride-on toy car with fully modeled interior components. Such scenes involve numerous interleaved objects with varied spatial and semantic relationships, leading to extreme geometric and semantic complexity.

However, due to this practical consideration, we chose not to focus our evaluation on such a setup, as texturing highly nested and composite scenes all at once is rare in typical workflows. In practice, these environments are more effectively and commonly constructed by texturing individual components with moderate levels of hidden internal structures (e.g., a closed toy chest, plush dolls, a bunk bed with built-in drawers, fully enclosed ride-on toy cars) separately, and then combining them into a larger scene graph or hierarchical assembly. Automating this compositional process remains a promising direction for future work.

---

Q2: A notable limitation of the method is that certain occluded areas on interior surfaces are neglected, as exemplified by the back of chairs on the bus shown in Figure 2.

A2: We thank the reviewer for the comment. We believe the concern may stem from the specific view shown in Figure 2 in the main paper, where only the front of the chairs is visible, and the backs are not shown from that particular camera angle. However, we would like to clarify that in our method, each mesh face is assigned a hit level based on multi-view ray casting from a densely sampled hemispherical camera rig (as mentioned in Section 4.1, L195). This setup ensures that all visible and occluded surfaces (including the backs of interior objects) are intersected by at least one ray and are thus included in the progressive texturing process.

---

Q3: The framework lacks a clear mechanism for adaptively adjusting the corresponding text prompt in response to varying hit levels.

A3: Thank you for bringing this up. Upon reviewing our submission, we acknowledge that we did not clearly describe how dual prompting is enabled within our framework and we appreciate the opportunity to clarify it.

As described in Section 3 (L172–174), we generate a texture for each hit level by rendering multi-view depth maps of the geometry corresponding to that level and synthesizing textures using a text-guided MVD module. Then, as noted in L178, the resulting textures from each hit level are merged into a unified UV texture map via visibility-weighted blending.

This design naturally enables layer-specific prompting: since the MVD module is conditioned on depth maps and a textual prompt independently per hit level, the prompt used at each level can be chosen independently. In practice, we assign a distinct prompt to the outermost layer (hit-level = 1) to control the exterior appearance, and a different prompt for hit levels ≥ 2 to control interior textures. This mechanism supports what we call dual prompting, as shown qualitatively in Figures 1 and 7.

We will make this sufficiently explicit in the revised version of the manuscript.

---

Q4: Potential issues in maintaining texture coherence when faces within a semantically coherent region are assigned to different hit levels

A4: Thank you for the thoughtful question.

As the reviewer correctly points out, our current hit level assignment is purely geometric, based on ray intersection depth, and therefore agnostic to semantic information. In certain edge cases, such as a vase with a narrow opening, semantically coherent interior regions (e.g., the connected inner surface) may be assigned to different hit levels due to partial visibility from external viewpoints. This could, in principle, result in textural discontinuities across those regions.

However, in practice, we empirically observe that our method still produces semantically plausible and coherent textures. We attribute this to several key factors:

1. Residual face clustering, combined with view-dependent normal flipping and backface culling (Section 3.2), preserves structural continuity across hit levels in the rendered depth maps. As shown in Figure 3 (right), although each hit level targets a distinct subset of faces (particularly those previously occluded), our method consistently maintains the overall geometry, retaining a globally coherent silhouette and shape in the depth maps. This inter-level consistency ensures that the diffusion model receives a stable geometric context at each synthesis stage, even though the specific regions being textured differ. As a result, as shown in the rendered textures in Figure 3 (right), semantic coherence is naturally preserved across hit levels.
2. In addition, our soft UV-space blending strategy (Section 3.3) further reinforces this continuity by smoothly interpolating textures across hit levels using view-dependent visibility weights. This weighted blending mitigates hard seams and ensures stylistic consistency in regions where semantic surfaces span multiple hit levels, as can be seen in Figure 4.

Together, these two components work in concert to ensure that our method produces coherent and semantically aligned textures, even in cases where semantically unified regions are geometrically split across hit levels.

That said, we acknowledge this limitation and view it as a promising direction for future work. Specifically, incorporating semantic-aware refinement of hit levels, for example by further grouping superfaces that belong to the same semantic volume with pre-trained part segmentation modules [*], could further improve consistency in such edge cases.

Additionally, while not explored in our current implementation due to computational resource constraints, an even more integrated strategy could involve performing soft UV-space blending during each denoising process itself, rather than after hit-level-specific texture synthesis. This would enable simultaneous multi-view and multi-hit-level texturing, potentially allowing for tighter coherence across depth layers.

We thank the reviewer for highlighting this nuanced and important point, and will clarify this limitation and potential mitigation strategies in the revision.

[*] Wu et al., Point Transformer V3: Simpler, Faster, Stronger, In CVPR, 2024

We sincerely thank the reviewer for their positive and encouraging feedback. We’re glad that the problem formulation and execution were found to be creative and solid, and that the rendered visuals were enjoyable and helpful.

---

Q1: It would be interesting to compare to NERF-like approaches that learn a full volumetric texture model rather than a uv-based texture model.

A1: We appreciate the reviewer’s suggestion to consider NeRF-like or volumetric texturing approaches such as FlashTex and Fantasia3D. As noted, in principle, these methods can support interior or cut-away views and can be combined with diffusion-based supervision (e.g., SDS loss) to optimize radiance fields for interior volumes or regions.

However, applying such methods directly without additional geometric and occlusion-aware constraints can be problematic. In practice, when training is performed primarily from a set of fixed external views, NeRF-like models tend to accumulate high density at the outermost surfaces. This effectively blocks camera rays from reaching the interior, preventing supervision signals from propagating to inner geometry. As a result, color radiance in these regions often remains under-constrained or unlearned.

One might attempt to mitigate this by gradually moving cameras inward or narrowing the frustum to expose internal volumes. However, without access to mesh structure, there is no guidance on how to move the cameras or which regions of the volume to expose. This often leads to indiscriminate optimization over the entire volume, introducing spurious geometry and severely degrading training efficiency. Even worse, NeRF-like models may recursively form nested high-density pseudo-surfaces, i.e. much like a troika structure, resulting in a degenerate, continuous shell-like volume that fails to capture meaningful interior surfaces.

On the other hand, if mesh structure is available, it becomes possible in principle to expose and supervise surface layers in a controlled, outside-in manner. But doing so effectively still requires an explicit mechanism to identify occluded geometry at each depth and to organize supervision accordingly. This is precisely the functionality that our method introduces through hit-level assignment and visibility-depth analysis.

Thus, we would like to emphasize that the core challenge in interior surface texturing is not the choice of representation (e.g., UV-based vs. volumetric), but the ability to perform geometry- and occlusion-aware visibility control. Once such control is established, the decision to optimize a UV texture map or a volumetric radiance field becomes secondary. Our framework provides the foundational mechanism required for either setting, and could be integrated into volumetric pipelines as a principled method for layered, surface-aware supervision, particularly in the presence of occluded or complex internal geometry.

---

Q2: Potential misleading use of "occlusion-aware texturing"

A2: We thank the reviewer for pointing out the potential ambiguity in the term “occlusion-aware.”

While occlusion is often treated as a view-dependent concept, it can also be view-invariant in the case of structurally enclosed interior regions. For instance, exterior surfaces may or may not be visible depending on the viewpoint, but interior surfaces that are fully enclosed by outer geometry remain occluded from all external views, regardless of camera position.

GOATex is specifically designed to address such occlusions. Our method explicitly identifies these deeply embedded, structurally occluded regions (i.e., faces that are never visible in standard multi-view setups) and progressively reveals them via ray-based visibility analysis. By decomposing the mesh into ordered hit levels and applying layer-wise texturing with blending, GOATex provides a geometric, structure-aware approach to handling occlusion that complements and extends conventional view-aware techniques.

In that sense, we believe the term “occlusion-aware” is appropriate, as our method does not rely solely on view-based visibility cues, but instead explicitly identifies and handles regions that are consistently hidden due to the mesh’s geometric enclosure. That said, we understand the concern and are open to adopting alternative terminology in future revisions to avoid confusion with view-dependent material modeling.