acceptodds
Withdrawn from ICLR 2027

CRAFT: Reference-Guided Surface-Addressed Texture Generation for Composable 3D Assets

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

While progress in native 3D texture generation can complete appearance beyond an observed view, existing texture generators often suffer from appearance bleeding when distinct geometric parts touch or intersect. This is because spatial voxelization can aggregate texture samples from multiple surfaces into a single texture state. Reference-image conditioning introduces a second ambiguity: features observed on a visible surface may be incorrectly transferred to nearby or occluded surfaces that share the same voxel. We introduce CRAFT, a native 3D texture-generation framework that augments each voxel coordinate with a discrete surface identity derived from UV-chart connectivity. This surface address preserves independent texture features for different surfaces throughout encoding, generation, and decoding. We further introduce visibility-routed reference conditioning, to establish accurate correspondence between visible surface tokens and reference-image features according to surface address and visibility. These designs preserve surface-level texture correspondence while retaining the efficiency of a sparse voxel-based representation. As a result, CRAFT reproduces the reference appearance more faithfully while reducing cross-surface color leakage and keeping distinct surface regions more clearly separated for component-level editing, separation, and recombination.

Primary area:
generative models
Keywords:
Native 3D generation, PBR material generation, 3D texture generation, sparse voxel representation, visibility-routed conditioning, multi-view image conditioning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bePUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CRAFT, a native 3D texture-generation framework that addresses intra-voxel or cross-surfaces texture bleeding between touching or intersecting surfaces by augmenting voxel coordinates with surface identities derived from a pre-provided UV-chart connectivity.

It further introduces visibility-routed reference conditioning, using first-hit visibility to associate the input reference-image features with the corresponding visible surfaces while preserving surface-specific texture representations.

Experiments demonstrate improved reference-view fidelity and occluded-region texturing, with support for surface separation, recombination, and image-guided texture editing.

Critical strengths

The practical motivation of the two new designs is valid: texture bleeding caused by limited spatial resolution and ambiguous surface identity. . The idea of preserving surface-specific texture features and routing reference-image features using visibility is intuitive and technical possible, despite for some major concerns as listed below.

Critical weaknesses

I have two major concerns over the fundamental assumptions of the work. 1. It relies on artist-created UV charts to define meaningful surface identities, which are often not available in typical AI-based 3D shape generation pipelines. 2. the first-hit visibility does not guarantee correct semantic correspondence between the reference images and the target geometry, and the results do not yet convincingly demonstrate practical effectiveness. More details can be found in below elaborations.

  1. Regarding the restrictive dependence on artist-created UV charts. The method assumes that the UV-chart induced connectivity is available to provide meaningful surface identities, either based on geometry feature or semantic feature. This is a significant restrictive assumption for current AI-based 3D shape generation pipelines using VAE+diffusion or flow matching. Typically, the texturing step often follows the 3D geometry (mesh or voxel) generation, for which artist-authored UV layouts are not be available. In fact , the obtained shape is often a monolithic unsegmented shape, and I suspect the method proposed in this paper does not work at all for automatically generated UV-chart such as via X-Atlas. Moreover, to segment a mesh to semantically meaningful isolated meshes required significant cognitive burden , once done, half of the task of 3D texturing is also completed.

  2. Establishing reference-image correspondences may be a very challenging task, especially for the scenarios targeted in this paper. For instance, the visibility-routing mechanism assumes that visible reference surfaces can be reliably associated with the target 3D mesh using calibrated cameras and first-hit visibility, despite that the reference images may depict different object configurations, viewpoints, or surface layouts. Since the target 3D scene can be significantly different (see the teaser image), visibility alone can not be easily detected, and moreover it does not guarantee semantic correspondence. As a result, the reported texture quality also appears insufficient and not convincing.

Overall, while the two proposed mechanisms are potentially promising, their relevance and effectiveness depends on very restrictive input assumptions and reliable surface correspondences; this dependence as well as its limited demonstrated texture quality raise substantial concerns about generality and practical applicability. To conclude, the contribution and quality of the paper is far from meeting the standard for ICLR.

Questions for authors

Some critical questions related to the above major weaknesses:

  1. How does CRAFT perform when artist-authored UV charts are unavailable, or when automatically generated UV charts do not align with meaningful geometric or semantic surface boundaries?

  2. Can your method handle monolithic meshes produced by standard 3D generation pipelines without manual UV authoring?

  3. How robust is visibility-routed conditioning when the reference images and target 3D geometry differ significantly in camera pose, part configuration, or semantic surface correspondence?

  4. 3D Native texture generation becomes a new trend, as you mentioned Why you did not compare your method with any of the STOA methods such as NaTex , LaFiTe and TEXTRIX , which also based on structured latent voxel-based native texture generation ?

4 comments
AuthorsOpenReview ↗

Author response to Concern #1: How does CRAFT perform when artist-authored UV charts are unavailable or do not align with meaningful surface boundaries?

We sincerely appreciate your insights and would like to address the concerns you raised regarding the practical effectiveness.

Concern #1: How does CRAFT perform when artist-authored UV charts are unavailable or do not align with meaningful surface boundaries?

Author response to Concern #1: We apologize that our original description gave the impression that CRAFT requires artist-authored UV charts at inference. We use UV-chart connectivity during training because it provides surface groups directly from textured meshes without additional annotations. At inference, CRAFT uses a surface ID to identify surfaces that should exchange texture information. These IDs can come from semantic labels, automatic part segmentation, or geometric partitioning.

In Appendix A, we evaluate five surface grouping strategies using the same trained model on 158 held-out PartObjaverse-Tiny assets. The table reports reconstruction in the reference view and completion on UV texels exposed by part separation.

Grouping strategyAvg. groups per assetReference-view PSNR ↑Reference-view SSIM ↑Reference-view LPIPS ↓Newly exposed UV PSNR ↑Newly exposed UV MSE ↓
Semantic part categories7.5826.1053420.9519280.03208913.810.05853
Individual part instances10.5826.0822540.9518100.03230613.520.06086
Geodesic partitioning11.9926.5876270.9535410.03171813.150.06520
Face-normal angle partitioning11.9126.3878890.9522740.03208013.070.06515
Merged XAtlas UV charts12.0026.00010.94940.03230812.440.09208

At approximately 12 groups per asset, completion PSNR for the three automatic strategies ranged from 12.44 to 13.15 dB. Reference-view quality was relatively stable across the five strategies at the tested granularity of roughly 8–12 groups per asset on average. Completion on newly exposed surfaces was more sensitive to the partition.

To address the reviewer’s question about automatically generated UV charts, we use XAtlas to generate charts directly from each mesh. We keep these charts fixed and vary how they are merged to obtain 12, 32, 64, 128, or 256 surface groups per asset. We evaluate reconstruction in the reference view and texture completion on UV texels corresponding to surfaces revealed by part separation.

XAtlas groups per assetReference-view PSNR ↑Reference-view SSIM ↑Newly exposed UV PSNR ↑Newly exposed UV MSE ↓
1226.00010.949412.440.09208
3226.07360.950512.290.09227
6426.11860.950412.120.09313
12826.05180.949811.870.09792
25626.12620.951611.680.10220

Reference-view PSNR varied by less than 0.13 dB across all five settings. Completion on newly exposed surfaces was more sensitive to fragmentation. As the group count increased from 12 to 256, UV-space PSNR fell at every step, from 12.44 to 11.68 dB, and MSE rose from 0.09208 to 0.10220.

These results show that CRAFT can use automatically generated XAtlas groups at inference, although fine partitions can separate an occluded region from the visible texture cues that guide its completion. We therefore recommend avoiding excessively fine partitions at inference, particularly when completing textures on occluded surfaces.

AuthorsOpenReview ↗

Author response to Concern #2: Can CRAFT handle monolithic meshes produced by standard 3D generation pipelines without manual UV authoring?

Concern #2. Can CRAFT handle monolithic meshes produced by standard 3D generation pipelines without manual UV authoring?

Author response to Concern #2: Yes. To address reviewer’s concern, we add a new Section 4.4 to illustrate our method can be integrated as a texture refinement module in standard 3D generation pipelines.

Specifically, we selected monolithic meshes from Tripo3D’s curated collection, used its part segmentation tool to identify semantic parts, and applied its part completion tool to close holes in the separated components. The original textures appeared coherent when the meshes were assembled. Separating the parts, however, revealed contact imprints at component interfaces and abrupt color changes on previously occluded surfaces (Fig. 7, Input Texture).

We illustrate CRAFT can be applied as a texture refinement module in the standard 3D generation pipelines. To refine the imprints and abrupt color, we render six views of the original assembled mesh to condition CRAFT and compare its results with those of other methods that accept multiple input views. The examples show that CRAFT fills the exposed contact regions with colors consistent with the surrounding texture on each part. We can also refine a selected part by holding the tokens of all other parts fixed, preserving their original high-resolution textures. Also, this setup required no manual UV authoring.

AuthorsOpenReview ↗

Author response to Concern #3: How robust is visibility-routed conditioning to differences in camera pose, part configuration, or semantic correspondence?

Concern #3. How robust is visibility-routed conditioning to differences in camera pose, part configuration, or semantic correspondence?

Author response to Concern #3. CRAFT addresses texture bleeding between contacting surfaces given an aligned reference image and a known camera pose. For each reference pixel, first-hit visibility identifies the target surface intersected by its camera ray and routes the image feature to that surface’s token. TEXGen uses a similar evaluation setup, rendering its conditioning image from a known viewpoint and passing the corresponding camera matrix to the texture generation model.

We apologize for the misleading impression our teaser may have given. We perform visibility routing and generate the texture with the parts in their original assembled configuration. We separate them only afterward to inspect the generated textures on previously occluded surfaces. The reference image and target mesh therefore share the same part configuration during visibility routing.

Section 4.4 illustrates how this setting arises in a modern 3D generation pipeline. We render the assembled mesh from six known viewpoints and use automatically predicted part IDs as surface groups. CRAFT then refines textures at contact regions while preserving the original textures of unselected parts.

As discussed in our response to Concern #1, we evaluated CRAFT with different surface groupings, including automatically generated partitions that do not necessarily align with semantic part boundaries. These experiments test how changes in surface grouping affect texture quality while the reference image and target mesh remain aligned.

Ablation under calibrated alignment. We evaluated the contribution of visibility routing and other components on the original 105-asset ablation set. The first table reports PSNR, SSIM, and LPIPS over the full reference-view rendering.

MethodPSNR ↑SSIM ↑LPIPS ↓
Ours27.1760.95670.0554
w/o surface-address packing25.0160.94830.0582
w/o BP. Ref. Injection24.7660.95000.0622
w/o visibility24.9320.94820.0643
w/o SA surface-isolation24.2110.94380.0632
w CA across surface23.9730.94580.0636

The second table reports PSNR, MAE, and RMSE on newly exposed surfaces, using masked pixels in the exploded view and masked texels in UV space.

MethodUV mask PSNR ↑UV mask MAE ↓UV mask RMSE ↓Explode mask PSNR ↑Explode mask MAE ↓Explode mask RMSE ↓
Ours19.9790.074700.1172518.4390.091260.14415
w/o surface-address packing16.7560.115380.1637316.1200.120720.17495
w/o BP. Ref. Injection15.5970.125810.1854014.9390.137650.20164
w/o visibility15.9740.121340.1777815.4640.129780.19068
w/o SA surface-isolation14.2910.151280.2089214.0320.155610.21578
w/ CA across surfaces14.9180.142460.2003814.5930.144880.20523

Removing visibility routing reduced reference-view PSNR from 27.176 to 24.932 dB. On newly exposed surfaces, PSNR fell from 18.439 to 15.464 dB in the exploded view and from 19.979 to 15.974 dB in UV space. These results support the contribution of visibility routing when the reference and target are aligned.

We also added Pixal3D as a baseline in the revised manuscript. CRAFT was inspired by Pixal3D’s approach to back-projecting reference-image features into a 3D representation. To control for camera pose, we provided Pixal3D with the same ground-truth reference camera pose used by CRAFT at inference. On the 158 held-out assets, CRAFT achieved a reference-view PSNR of 26.105 dB and LPIPS of 0.0321, compared with 25.726 dB and 0.0365 for Pixal3D. After part separation, Figure 4 shows that moving a white book reveals a white imprint on the wooden tabletop in the Pixal3D result. CRAFT reduces this imprint, illustrating its ability to limit texture bleeding between contacting surfaces.

If the reference image and target mesh depict different part configurations, first-hit visibility alone cannot establish semantic correspondence. For example, if a book appears in different positions in the reference image and target mesh, a pixel depicting the book may project onto the tabletop, causing its feature to be routed to the table.

AuthorsOpenReview ↗

Author response to Concern #4: 3D Native texture generation becomes a new trend, as you mentioned Why you did not compare your method with any of the STOA methods such as NaTex , LaFiTe and TEXTRIX , which also based on structured latent voxel-based native texture generation?

Concern #4: 3D Native texture generation becomes a new trend, as you mentioned Why you did not compare your method with any of the STOA methods such as NaTex , LaFiTe and TEXTRIX , which also based on structured latent voxel-based native texture generation ?

Author response to Concern #4:We could not evaluate NaTex, LaFiTe, and TEXTRIX because its code and checkpoint were unavailable. Direct quantitative comparisons with these three methods therefore remain open.

Our evaluation includes TRELLIS.2, and the revision adds both quantitative and qualitative comparisons with Pixal3D. These two methods are also voxel-based native texture generation methods. Pixal3D is particularly relevant because it also back-projects image features into a 3D representation.

As mentioned in the response to Concern #3, we provided Pixal3D with the same ground-truth reference camera pose used by CRAFT at inference. On the 158 held-out assets, reference-view PSNR was 26.105 dB for CRAFT, 25.726 dB for Pixal3D, and 22.014 dB for TRELLIS.2. On newly exposed pixels in the exploded view, the corresponding scores were 14.362, 14.283, and 11.611 dB. In the revised manuscript, Figure 4 in the revised manuscript further illustrates the practical difference at a contact surface: moving a white book reveals a white imprint on the wooden tabletop in the Pixal3D result, while CRAFT reduces the imprint.

We made every effort to include all relevant native 3D texture generation methods for which we could obtain runnable inference code and checkpoints. We hope the comparisons with TRELLIS.2 and Pixal3D provide a convincing evaluation of our work.

Reviewer_GZWCOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

CRAFT reduces texture bleeding between nearby surfaces by augmenting voxel coordinates with UV-chart identities and routing reference-image features to the first visible surface. It builds on TRELLIS.2, preserving separate surface states through encoding, generation, and decoding. The approach is practical, and the visual results are encouraging. However, some issues limit my confidence in the main conclusions.

Critical strengths
  • The paper identifies a concrete limitation of voxel-based texturing and clearly distinguishes representation mixing from incorrect reference-feature assignment.
  • The method reuses a pretrained backbone through a simple, coherent design.
  • The separated views, particularly the reconstruction and generation comparisons in Figure 6, provide convincing examples of reduced contamination. The reported quantitative gains are also meaningful.
Critical weaknesses
  • The occluded-region evaluation is unclear. Appendix E.1 describes separating components but does not state whether metrics are restricted to surfaces that were invisible in the reference and become visible after separation. Exploded renderings also contain previously visible surfaces and background. If evaluated as whole images, gains could partly reflect better visible-surface reconstruction rather than better hidden-surface completion. This ambiguity affects a central claim.

  • The experiments do not adequately test the surface-partition design. Appendix A appears to vary UV-island layout rather than chart connectivity. It therefore does not establish robustness to different seams or partitions. Strictly isolating charts may impair continuity across seams, while self-contact within one chart can retain the original ambiguity. Moreover, removing surface packing changes several factors at once. A comparison that preserves surface addresses but allows cross-surface attention would better establish whether strict isolation is necessary.

  • Training–test independence is not established. Training uses TexVerse and testing uses Objaverse-XL, but no asset-level deduplication procedure is reported. TexVerse documents overlap with Objaverse, so using different dataset names is insufficient. This does not demonstrate leakage in the selected splits, but source-ID and duplicate checks are needed to support the generalization results.

Questions for authors
  1. Are occluded-region metrics computed only on newly visible surfaces? Please specify the mask and metric aggregation, and provide results on this region if the current evaluation uses full renderings.
  2. Does Appendix A change chart connectivity or only island arrangement? If connectivity is unchanged, what evidence supports robustness to alternative chart partitions, particularly for texture continuity across seams?
  3. Were the 20k training assets and 105 test assets checked for shared source IDs, including different versions of the same asset? Please describe the checks and report any overlap found.
7 comments
AuthorsOpenReview ↗

Author response to Concern #1: Were the 20k training assets and 105 test assets checked for shared source IDs, including different versions of the same asset? Please describe the checks and report any overlap found.

We sincerely appreciate the effort to review the previous manuscript and the opportunity to address your concerns. We provide our point-by-point responses below.

Concern #1: Were the 20k training assets and 105 test assets checked for shared source IDs, including different versions of the same asset? Please describe the checks and report any overlap found.

Author response to Concern #1: We appreciate your careful attention to the overlap between the training and test sets. We apologize for not performing this check before submission. We have now audited all 105 original test assets against the 16,442 assets used for training. We confirmed 52 overlaps by source ID and identified 28 additional test assets as possible versions of training assets. In total, 80 of the 105 test assets had a confirmed or potential match. Given this overlap, we re-evaluated our method on a separately screened test set. We describe the checks and the revised results below.

Source-ID matching. We collected the filenames from the TexVerse-1K training directory and excluded assets whose sequences exceeded 32,768 tokens and caused out-of-memory errors. These assets did not enter training, leaving 16,442 training IDs. For each training asset, we used the glb_paths field to link its local filename to its TexVerse-1K metadata entry. We retrieved the 32-character source ID from the entry key and the Sketchfab model UID from its thumbnail URL. On the test side, we verified that the filenames corresponded to Objaverse-XL source IDs. We compared each test ID against all three training identifiers: the metadata source ID, the Sketchfab UID, and the filename-derived ID. This identified 52 test assets with matching identifiers.

Screening for different versions of the same asset. To detect identical meshes stored under different IDs, we hashed the vertices and triangle faces of each normalized mesh after removing duplicates and sorting both into a canonical order. One test asset had the same geometry as three training assets. Visual inspection showed that its texture differed from all three. We report this as an exact geometry match and exclude it from the textured-asset duplicate count. A rotated re-upload can produce a different mesh hash. To screen for such cases, we sampled points on each mesh in proportion to surface area and compared rotation-invariant shape descriptors: PCA eigenvalues, radial-distance quantiles, and quantiles of pairwise point distances (D2). This flagged 28 additional test assets. Geometric similarity alone does not establish that two assets share a source, so we report these separately from the 52 confirmed ID matches.

Revised evaluation. Given this overlap, we recognize that the original evaluation does not reliably measure generalization to unseen assets. We then downloaded and applied the same checks to all 200 assets in PartObjaverse-Tiny, which is also a subset of Objaverse-XL. We excluded six assets flagged by the overlap checks and a further 36 that were unsuitable for texture evaluation because they were untextured white models, used alpha transparency, or yielded abnormally dark conditioning images with the TRELLIS.2 data toolkit. This left 158 test assets with no detected training-set overlap. The table below reports single-view texture generation results on the full reference-view rendering, including the background. Following the suggestions of Reviewers d1AJ and bePU, we also included Pixal3D.

Table 1. Quantitative comparison on the full reference view.

MethodPSNR ↑SSIM ↑LPIPS ↓
MaterialMVP23.5810810.9378970.044907
TexGen24.5882420.9353440.056210
TRELLIS.222.0138690.9277260.056387
UniTEX23.2050110.9378690.048676
Pixal3D25.7258420.9494200.036529
Ours26.1053420.9519280.032089

On the new test set, our method achieved the highest PSNR, SSIM, and lowest LPIPS among the six methods. In the revised manuscript, we update the description of training and test data in Section 4.1. We document the overlap checks and test-set exclusions in Appendix I. Also, we report the results on the 158 retained assets in Table 1.

AuthorsOpenReview ↗

Author response to Concern #2: Are occluded-region metrics computed only on newly visible surfaces? Please specify the mask and metric aggregation, and provide results on this region if the current evaluation uses full renderings.

Concern #2: Are occluded-region metrics computed only on newly visible surfaces? Please specify the mask and metric aggregation, and provide results on this region if the current evaluation uses full renderings.

Author response to Concern #2: Previously, the occluded-region metrics were not computed exclusively on newly visible surfaces. We recognize that these scores combined visible-surface reconstruction and hidden-surface completion. The revised evaluation measures only surfaces that were occluded in the reference view and became visible after component separation.

We evaluate newly visible surfaces in image space. As suggested by Reviewer d1AJ, we also compute the metrics directly on the resulting uv-space texture. For each foreground pixel in the exploded rendering, we use its triangle index and barycentric coordinates to locate the corresponding point on the assembled mesh. We retain the pixel if that point was occluded in the reference view. In UV space, we select texels on newly exposed surfaces and retain those with valid surface correspondence across all methods. Within each evaluation, every method uses the same mask for a given asset.

We report PSNR, MAE, and RMSE because they let us measure color errors directly within these masks. Our original CLIP-based metrics summarize the entire image, and LPIPS features can depend on pixels outside the mask. For each asset, we compute all three metrics from sRGB values normalized to [0,1][0,1], giving equal weight to each selected pixel or texel. We then take the arithmetic mean of the per-asset scores across the 158 test assets, so each asset contributes equally regardless of mask size.

Results on newly exposed surfaces across 158 test assets.

MethodExploded view: PSNR ↑MAE ↓RMSE ↓UV space: PSNR ↑MAE ↓RMSE ↓
Ours14.3620.157930.2105613.8120.192060.24195
Pixal3D14.2830.160630.2180513.5130.192600.24411
TRELLIS.211.6110.247710.3061111.0250.297890.34215
TexGen11.3210.227630.3055710.4070.279300.34546
MaterialMVP9.9720.248380.3311410.7810.241550.30685
UniTEX9.1520.275250.364889.7700.270830.34236

Among the six methods, our method achieved the highest PSNR and the lowest MAE and RMSE in both the image space and the UV space. Because artists created the ground-truth textures, these comparisons give some indication of how closely our method reflects the choices artists make when texturing surfaces hidden from the reference view.

In the revised manuscript, Table 1 reports the image-space and UV-space results for the newly exposed surfaces. Appendix H.1 describes the evaluation masks, metrics, and aggregation procedure.

AuthorsOpenReview ↗

Author response to Concern #3: Does Appendix A change chart connectivity or only island arrangement? If connectivity is unchanged, what evidence supports robustness to alternative chart partitions, particularly for texture continuity across seams?

Concern #3: Does Appendix A change chart connectivity or only island arrangement? If connectivity is unchanged, what evidence supports robustness to alternative chart partitions, particularly for texture continuity across seams?

Author response to Concern #3: We apologize for the ambiguity in our original description. The experiment in the original Appendix A, now Appendix B, varies the arrangement of UV islands in a single texture atlas while keeping the chart connectivity used by CRAFT fixed. Given an input mesh, we first assign a chart ID to each isolated UV chart based on the original UV connectivity. We then use xatlas to pack the mesh into a single texture atlas and sample points on the textured surface. Each point retains its original chart ID, and we aggregate points into tokens indexed by (chart ID, 3D voxel position). Across the layouts shown in Appendix B, the islands may be re-organized by xatals vary. However, the same surface points retain the same chart IDs. The original experiment therefore shows that CRAFT produces similar textures under different atlas layouts.

Surface partitions robustness. In the revised Appendix A, we test the partition itself by changing which surface points share an ID at inference. We use the same trained model and the same 158 held-out PartObjaverse-Tiny assets throughout. We compare five grouping strategies and report texture quality both in the full reference view and on newly exposed surfaces in UV space.

Grouping strategyAvg. groups per assetReference-view PSNR ↑Reference-view SSIM ↑Reference-view LPIPS ↓Newly exposed UV PSNR ↑Newly exposed UV MSE ↓
Semantic part categories7.5826.1053420.9519280.03208913.810.05853
Individual part instances10.5826.0822540.9518100.03230613.520.06086
Geodesic partitioning11.9926.5876270.9535410.03171813.150.06520
Face-normal angle partitioning11.9126.3878890.9522740.03208013.070.06515
Merged XAtlas UV charts12.0026.00010.94940.03230812.440.09208

Reference-view PSNR varied by less than 0.6 dB across all five strategies, which averaged roughly 8–12 groups per asset. The test assets have approximately 11 annotated part instances on average, so we targeted 12 groups for each of the three automatic methods to compare different boundary placements at a similar group count. Across these three methods, UV-space PSNR on newly exposed surfaces ranged from 12.44 to 13.15 dB. These results show that a model trained with UV-chart groups can use alternative partitions at inference. Section 4.4 provides a qualitative example on generated meshes: using part IDs predicted by Tripo3D, CRAFT fills previously occluded contact regions with colors consistent with the surrounding texture on each part.

Texture continuity across seams. We address this concern indirectly by testing what happens when more chart seams become boundaries between groups. We keep the underlying XAtlas charts fixed and vary how they are merged to form 12 to 256 groups per asset. Charts merged into the same group can exchange texture cues across their shared seam. When the charts belong to separate groups, that seam also becomes a boundary for within-group token interactions. We measure the effect on texture completion using PSNR and MSE over newly exposed UV texels.

XAtlas groups per assetReference-view PSNR ↑Reference-view SSIM ↑Newly exposed UV PSNR ↑Newly exposed UV MSE ↓
1226.00010.949412.440.09208
3226.07360.950512.290.09227
6426.11860.950412.120.09313
12826.05180.949811.870.09792
25626.12620.951611.680.10220

As the number of groups increased, reference-view PSNR varied by less than 0.13 dB. On newly exposed surfaces, however, UV-space PSNR fell at every step, from 12.44 to 11.68 dB, while MSE rose from 0.09208 to 0.10220. This trend is consistent with finer partitions separating occluded regions from visible texture cues on the same continuous surface. The experiment thus identifies a cost of excessive fragmentation for hidden-surface completion. Continuity across individual seams remains unquantified, since these metrics average errors over newly exposed texels and do not directly compare adjacent surface points on opposite sides of a seam.

In the revised manuscript, we clarify in Appendix B that the original experiment varies UV-island placement while preserving chart connectivity. We add a new section in Appendix A that compares five surface grouping strategies and examines how increasing the number of XAtlas groups affects hidden-surface completion, providing indirect evidence about texture propagation across chart seams. Section 4.4 further demonstrates texture refinement using automatically predicted part IDs.

Reviewer_d1AJOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CRAFT, a reference-guided 3D texture generation framework designed to reduce texture bleeding between spatially adjacent, touching, or intersecting surfaces. The authors identify two sources of ambiguity in voxel-based 3D texture generation: (1) intra-voxel texture entanglement, where distinct surfaces are aggregated into the same voxel representation, and (2) surface-ambiguous reference conditioning, where image features are incorrectly assigned to neighboring or occluded surfaces.

To address these issues, CRAFT introduces a surface-indexed representation that augments voxel coordinates with discrete surface identities derived from UV-chart connectivity. The method preserves separate appearance features through VAE encoding, diffusion-based generation, and decoding. It additionally introduces visibility-routed reference image conditioning, using first-hit surface correspondences to inject image features into the appropriate visible surface tokens.

The approach builds on a pretrained TRELLIS.2 model and fine-tunes its material generation network. Experiments on the Objaverse-XL dataset demonstrate improvements in reference-view reconstruction and occluded-region texture generation compared with several existing methods. Qualitative results further illustrate component separation and image-guided texture editing.

Critical strengths
  1. Well-motivated methodology. The paper introduces two complementary designs: (1) UV-chart connectivity-based surface partitioning to prevent intra-voxel texture entanglement, and (2) surface-aware pixel back-projection for accurate reference-image conditioning.
  2. Promising results. The method demonstrates noticeable improvements in both visible-view reconstruction and occluded-region generation.
Critical weaknesses
  1. Lack of quantitative validation of object texture disentanglement. Reducing cross-surface texture bleeding is a central contribution, yet the paper primarily relies on qualitative comparisons and general image-quality metrics. Direct measurements of cross-surface contamination would better support this claim.
  2. Incomplete ablation studies. Table 2 appears to use a different evaluation setting from Table 1, making it difficult to attribute the reported improvements to surface addressing and visibility routing. Separate ablations for visible-view reconstruction and occluded-region generation, consistent with Table 1, are needed to clarify each component's contribution.
  3. Limited evaluation of composability and editing. These capabilities are demonstrated primarily through qualitative examples, without quantitative evidence of their benefits.
Questions for authors
  1. Why are Pixal3D and TEXTRIX not included in the comparisons, given their relevance to the proposed method?
  2. How reliably does UV-chart connectivity correspond to meaningful surface groups? What are the potential failure cases, and how frequently do they occur?

Suggestions:

  1. Provide quantitative measurements of cross-surface texture leakage to strengthen the central claim. For example: UV-space texture color differences.
  2. Include ablation studies for both visible-view reconstruction and occluded-region generation using consistent evaluation protocols.

I appreciate the well-motivated and elegant approach to addressing intra-voxel texture entanglement. Addressing these concerns, particularly through direct measurements of texture leakage and more comprehensive ablations, would substantially strengthen the paper's soundness. I would be inclined to increase my rating if these concerns are adequately addressed.

4 comments
AuthorsOpenReview ↗

Author response to Concern #1: Why are Pixal3D and TEXTRIX not included in the comparisons, given their relevance to the proposed method?

We sincerely appreciate the effort to review the previous manuscript and the opportunity for us to address your concern.

Concern #1 : Why are Pixal3D and TEXTRIX not included in the comparisons, given their relevance to the proposed method?

Author response to Concern #1 : In the revised manuscript, we have added qualitative and quantitative comparisons with Pixal3D. We could not evaluate TEXTRIX because its code and checkpoint were unavailable.

Qualitative comparison with Pixal3D. Figure 4 now includes a Pixal3D column for single-view texture generation. As shown in the figure, separating the parts of Pixal3D generation assets reveals artifacts where surfaces touch or intersect. For example, moving the white book exposes a white imprint on the wooden table where the book rested. This is because Pixal3D, as a voxel-based method, uses a single texture state per voxel, introducing intra-voxel texture entanglement. In contrast, our method effectively suppresses these artifacts.

Quantitative comparison with Pixal3D on the full reference view. Following Reviewer GZWC’s suggestion, we checked the original benchmark for overlap with our training set and found that some test assets had also appeared in training. This limits the conclusions we can draw from the original evaluation about performance on unseen assets. We therefore repeated the comparison on PartObjaverse-Tiny. We excluded six assets flagged by the overlap checks and 36 untextured white models, assets with alpha transparency, or assets that produced abnormally dark conditioning images. This left 158 test assets. The table reports single-view texture generation results on the full reference-view rendering, including the background. Our method achieved the highest PSNR, SSIM and lowest LPIPS among the six methods.

Table 1. Quantitative comparison on the full reference view.

MethodPSNR ↑SSIM ↑LPIPS ↓
MaterialMVP23.5810810.9378970.044907
TexGen24.5882420.9353440.056210
TRELLIS.222.0138690.9277260.056387
UniTEX23.2050110.9378690.048676
Pixal3D25.7258420.9494200.036529
Ours26.1053420.9519280.032089

Quantitative comparison with Pixal3D on newly exposed surfaces. Reviewer GZWC also suggested restricting the occluded-region evaluation to surfaces that become visible when the parts are separated, so that any quantitative gains can precisely reflect improvements in hidden-surface completion rather than in visible-surface reconstruction. Thus, we construct a mask of these newly exposed surfaces and compute the metrics only within the masked region. Our original evaluation computed LPIPS, CLIP, CLIP-FID, and CMMD on full renderings. The CLIP-based metrics use image-level features, and the receptive fields of LPIPS features can extend across the mask boundary. These metrics are therefore less suitable for evaluating the masked region in isolation. In the revised manuscript, we report PSNR, MAE, and RMSE computed from color differences only at the selected pixels in image space. The table below reports the single-view results on newly exposed surfaces after part separation.

Table 2. Comparison of single-view texture generation on newly exposed surfaces after part separation.

MethodPSNR ↑MAE ↓RMSE ↓
Ours14.3620.157930.21056
Pixal3D14.2830.160630.21805
TRELLIS.211.6110.247710.30611
TexGen11.3210.227630.30557
MaterialMVP9.9720.248380.33114
UniTEX9.1520.275250.36488

Across the 158 assets, our method achieved the highest PSNR and the lowest MAE and RMSE. Because the ground-truth textures were created by artists, these results reflect, to some extent, how closely our method matches the artists’ texture choices on surfaces hidden from the reference view. Table 1 in the revised manuscript reports the full reference-view results and the masked image-space and UV-space results on newly exposed surfaces. Appendix H.1 describes the evaluation masks, metrics, and aggregation across assets.

AuthorsOpenReview ↗

Author response to Concern #2: Provide quantitative measurements of cross-surface texture leakage to strengthen the central claim. For example: UV-space texture color differences.

Concern #2: Provide quantitative measurements of cross-surface texture leakage to strengthen the central claim. For example: UV-space texture color differences.

Author response to Concern #2: In the revised manuscript, we have added a UV-space evaluation that measures color differences between the generated and ground-truth textures. We first compute PSNR, MAE, and RMSE over texels within UV islands that have valid surface correspondence across all methods, excluding the empty space outside the islands.

Table 3. UV-space texture comparison over the full atlas on 158 test assets.

MethodPSNR ↑MAE ↓RMSE ↓
Ours13.730.184460.23893
Pixal3D13.620.185690.23896
MaterialMVP11.140.223700.29132
TRELLIS.210.990.290130.33926
TexGen10.730.257100.32893
UniTEX10.470.243590.31546

For the same consideration discussed in our response to Concern #1, we restrict the comparison to texels whose corresponding surface points were hidden in the reference view and became visible after part separation. This mask isolates color differences on newly exposed surfaces, including contact regions where cross-surface texture leakage can occur.

Table 4. UV-space texture comparison on newly exposed surfaces.

MethodPSNR ↑MAE ↓RMSE ↓
Ours13.8120.192060.24195
Pixal3D13.5130.192600.24411
TRELLIS.211.0250.297890.34215
MaterialMVP10.7810.241550.30685
TexGen10.4070.279300.34546
UniTEX9.7700.270830.34236

Our method achieved the highest PSNR and the lowest MAE and RMSE in both evaluations. After excluding surfaces visible in the reference view, our method still achieved higher PSNR, along with lowest MAE and RMSE. These results show that our gains extend beyond visible-surface reconstruction to texture completion on newly exposed surfaces. The UV-space scores measure texture error on newly exposed surfaces.

Figure 4 in the revised manuscript shows where colors or patterns transfer between contacting surfaces. Table 1 in the revised manuscript includes the masked image-space and UV-space results shown in Tables 2 and 4 of this response. Appendix C reports the full-atlas results shown in Table 3 of this response.

AuthorsOpenReview ↗

Author response to Concern #3: Include ablation studies for both visible-view reconstruction and occluded-region generation using consistent evaluation protocols.

Concern #3: Include ablation studies for both visible-view reconstruction and occluded-region generation using consistent evaluation protocols.

Author response to Concern #3: In the original manuscript, we reported results for the exploded view without evaluating the normal render view or masking the newly exposed surfaces. In the revised manuscript, to be consistent with the evaluation above, we evaluate visible-view reconstruction on the full rendering and report PSNR, SSIM, and LPIPS. We also evaluate the newly exposed surfaces using masks in both the exploded view and UV space and report PSNR, MAE, and RMSE. The results are shown below.

Ablation study on visible view

MethodPSNR ↑SSIM ↑LPIPS ↓
w/o surface-address packing25.0160.94830.0582
w/o BP. Ref. Injection24.7660.95000.0622
w/o visibility24.9320.94820.0643
w/o SA surface-isolation24.2110.94380.0632
w/ CA across surfaces23.9730.94580.0636
Ours27.1760.95670.0554

Ablation study on texture completion for newly exposed surfaces.

MethodImage-space PSNR ↑Image-space MAE ↓Image-space RMSE ↓UV-space PSNR ↑UV-space MAE ↓UV-space RMSE ↓
w/o surface-address packing16.1200.120720.1749516.7560.115380.16373
w/o BP. Ref. Injection14.9390.137650.2016415.5970.125810.18540
w/o visibility15.4640.129780.1906815.9740.121340.17778
w/o SA surface-isolation14.0320.155610.2157814.2910.151280.20892
w/ CA across surfaces14.5930.144880.2052314.9180.142460.20038
Ours18.4390.091260.1441519.9790.074700.11725

Following Reviewer GZWC’s suggestion, we added two ablations to test whether the DiT benefits from isolating tokens by surface address. We compare both variants with w/o BP. Ref. Injection because all three inject image features only through cross-attention.

  • w/o SA surface-isolation retains surface addresses but removes the self-attention isolation mask. Each token can attend to tokens from all surfaces of the same object, including its own surface.
  • w/ CA across surfaces retains surface-isolated self-attention and adds a cross-attention block afterward. Tokens from the current surface serve as queries. Only tokens from the other surfaces of the same object serve as keys and values.

Both variants perform worse than w/o BP. Ref. Injection on every reported metric. On newly exposed surfaces, removing isolation reduces UV-space PSNR from 15.597 to 14.291 dB, while adding cross-surface attention reduces it to 14.918 dB. Qualitatively, unrestricted self-attention produces pronounced color bleeding between surfaces. The cross-surface attention variant shows less bleeding, but image details are sometimes assigned to the wrong surface.

Section 4.3 discusses the generation ablations, Table 2 reports their quantitative results, and Figure 5 provides qualitative comparisons of ablations.

AuthorsOpenReview ↗

Author response to Concern #4: How reliably does UV-chart connectivity correspond to meaningful surface groups? What are the potential failure cases, and how frequently do they occur?

Concern #4: How reliably does UV-chart connectivity correspond to meaningful surface groups? What are the potential failure cases, and how frequently do they occur?

Author response to Concern #4: Two kinds of mismatch between UV-chart connectivity and meaningful surface groups can occur. A UV seam can split a continuous part, separating an occluded region from the visible texture cues that would guide its completion. Conversely, a single chart can span an appearance boundary, allowing texture cues to pass between regions that should remain distinct. We did not annotate these mismatches asset by asset, so we cannot reliably estimate how often either case occurs.

CRAFT uses UV-chart connectivity as one practical way to define surface groups. We used it during training and in our original evaluation because it provides groups directly from textured meshes. In the revised evaluation, we applied the same trained model to 158 PartObjaverse-Tiny assets using their annotated semantic part labels as surface IDs.

To test how well CRAFT transfers to alternative surface groupings, we compared five strategies on these same assets.

Grouping strategyAvg. groups per assetReference-view PSNR ↑Reference-view SSIM ↑Reference-view LPIPS ↓Newly exposed UV PSNR ↑Newly exposed UV MSE ↓
Semantic part categories7.5826.1053420.9519280.03208913.810.05853
Individual part instances10.5826.0822540.9518100.03230613.520.06086
Geodesic partitioning11.9926.5876270.9535410.03171813.150.06520
Face-normal angle partitioning11.9126.3878890.9522740.03208013.070.06515
Merged XAtlas UV charts12.0026.00010.94940.03230812.440.09208

Semantic groups surfaces by their annotated part category, while Instance assigns a separate group to each annotated part instance. The test assets have approximately 11 annotated part instances on average. We therefore targeted 12 groups for each of the three automatic methods to compare different ways of partitioning a surface at a similar part count. Geodesic recursively divides the surface into regions of similar area using distances along the mesh. Angle splits the mesh where adjacent face normals differ by at least 45° and merges small regions. XAtlas generates UV charts from the geometry and merges whole charts into groups.

Reference-view quality was relatively stable across all five strategies, which averaged roughly 8–12 groups per asset. At approximately 12 groups per asset, UV-space PSNR on newly exposed surfaces ranged from 12.44 to 13.15 dB across the three automatic strategies. These results suggest that when UV-chart boundaries do not align with meaningful surface groups, CRAFT can still work with automatically generated partitions at the tested granularity.

These results also show that CRAFT can use semantic groups at inference even though it was trained with UV-chart groups. Section 4.4 provides a complementary qualitative example. Using part IDs predicted by Tripo3D, CRAFT fills previously occluded contact regions of generated 3D model with colors consistent with the surrounding texture on each part.

One potential failure case is excessive fragmentation, where a continuous surface is split into so many groups that an occluded region can lose access to the visible texture cues needed for its completion. To test this possibility, we kept the XAtlas grouping method fixed and increased the number of groups per asset from 12 to 256.

XAtlas groups per assetReference-view PSNR ↑Reference-view SSIM ↑Newly exposed UV PSNR ↑Newly exposed UV MSE ↓
1226.00010.949412.440.09208
3226.07360.950512.290.09227
6426.11860.950412.120.09313
12826.05180.949811.870.09792
25626.12620.951611.680.10220

The more groups we divided each asset into, the lower the average UV-space completion PSNR on newly exposed surfaces. It fell at every step, from 12.44 dB with 12 groups to 11.68 dB with 256 groups. Over the same range, reference-view PSNR varied by less than 0.13 dB. This trend is consistent with the way CRAFT propagates texture cues within groups. Finer partitions can separate an occluded region from the visible surface that would otherwise guide its completion.

In the revised manuscript, we clarify how UV charts define surface groups during training. We add a new section in Appendix A reporting experiments on how the grouping strategy and the number of surface groups affect texture generation. We also add a new section (Section 4.4) presenting an experiment that demonstrates how automatically segmented parts can guide texture refinement at inference.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.