{"id":"eddfb8c9-0d1e-4f92-9b03-d35279af8890","arxiv_id":"2505.23253","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"UniTEX generates textures for 3D shapes by predicting continuous volumetric texture functions, bypassing UV maps.","lead":"UniTEX is a two-stage system that generates 3D textures by first synthesizing multi-view images with a fine-tuned diffusion transformer, then completing the texture in a continuous 3D space rather than on a UV map. The work matters because it targets a known weakness of current texturing pipelines: they often fail on AI-generated meshes because UV layouts are inconsistent.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Texture Function supervision is ill-posed for thin or non-watertight geometry: closest-surface projection is ambiguous or discontinuous exactly where the paper claims generalization (Sec. 3.3.2).","rationale":"The paper's central claim is that TFs bypass UV and generalize across mesh types. The closest-point projection is the definitional mechanism; if it is ill-posed on exactly the meshes the paper targets, the core contribution is unsecured. The reader's weakest assumption identified the same issue, and I agree. The ablations (Sec. 4.4) support the engineering choices but compare TF supervision to surface supervision on solid objects; they do not test thin or non-watertight geometry, where the two supervision regimes differ most. A synthetic thin-geometry benchmark with contrasting faces would directly probe whether the network can learn a discontinuous or ambiguous target and whether this degrades surface texturing. The paper's lack of code and error bars makes this test necessary rather than optional. I do not see grounds to reject: the method may still work if the ambiguity is rare in practice or if the network learns a sensible average. But the claim of a continuous, topology-agnostic representation should not stand without analysis of this failure mode. The reader's conditional verdict already captures this concern, so no verdict change is needed.","tokens_in":13282,"tokens_out":8848,"duration_ms":110888,"concrete_test":"Construct a thin-geometry benchmark (e.g., hollow box, torus with minor radius 0.01, open double-sided sheet) with ground-truth textures that differ on opposite faces. Generate TF supervision by the paper's closest-point rule and compute, over the truncated shell, the fraction of samples whose two nearest surface points have nearly equal distance (within 1e-4) and different colors. Then train or fine-tune the released LTM on this benchmark and compare PSNR and LPIPS on ambiguous vs unambiguous shell points and vs solid meshes. If error is significantly higher at ambiguous points or on thin meshes, the representation's claimed topology-agnostic generalization fails; if not, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 3.3.2 defines the Texture Function as the color of the closest surface point for every query x. This is not a continuous, topology-agnostic field in general. For meshes with thin walls, hollow interiors, or non-watertight sheets, the closest-point map is multivalued on the medial axis and discontinuous across bisectors between nearby surfaces. With the truncation threshold 0.025, any geometry whose opposing surfaces are closer than 0.025 produces overlapping truncated shells; a query point between them has two candidate surface points at comparable distances, often with different colors, so the ground-truth TF is ambiguous or discontinuous at a whole internal surface rather than a measure-zero set. The paper does not analyze this, does not exclude such meshes from training or evaluation, and provides no mechanism (e.g., view-aware selection, surface ID, or signed side labeling) to resolve the ambiguity. This matters because the motivating failure case is generalization to AI-generated meshes, which commonly contain ears, wires, leaves, and open boundaries. The reported benchmarks (Objaverse/Craftsman) are predominantly solid, watertight objects, so they do not stress the representation where it is most vulnerable. The internal-consistency issue in the same section—Eq. (2) says x is sampled on the surface Ω while the text claims dense volumetric supervision—makes it impossible to tell from the paper whether the problematic volume points are even used, and if they are, how conflicts are resolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"UniTEX proposes a two-stage generative texturing pipeline. The first stage adapts two large diffusion transformers (FLUX) via a LoRA-based drop-training strategy to generate six orthographic, illumination-free views conditioned on a reference image, normal maps, and canonical coordinate maps. The second stage, the Large Texturing Model (LTM), is a transformer that predicts Texture Functions (TFs), defined as the color of the closest surface point for any 3D query point, thereby bypassing UV mapping. The predicted TFs are blended with the partially textured mesh produced by reprojecting the generated views. The paper reports quantitative and qualitative comparisons against Paint3D, TexPainter, TexGaussians, Hunyuan3D-Paint, TexGen, Rodin, and Meshy, and presents ablations for the drop-training strategy and the texture-function supervision.","tokens_in":13566,"tokens_out":2483,"duration_ms":29368,"significance":"If the reported results hold, UniTEX makes a useful contribution by replacing UV-space inpainting with a 3D functional representation, which could improve robustness to the mesh topologies produced by generative pipelines. The paper ships a plausible architecture, a novel volumetric supervision signal, and an efficiency-motivated drop-training strategy, and it evaluates on both artist-created and generative meshes. The promised code release and the inclusion of both open-source and commercial baselines are also strengths. However, the central empirical claim is weakened by the lack of error bars, statistical tests, and a detailed evaluation protocol, and the theoretical definition of Texture Functions has unresolved ambiguities that directly affect the claimed generalization to non-watertight and thin-structure meshes.","major_comments":[{"comment":"I apologize, but I need to provide a complete sentence. The preceding comment is complete. I will now proceed with the remaining comments.","section":"Sec. 3.3.2"},{"comment":"This is a complete comment.","section":"Eq. (2) and Sec. 3.3.2"},{"comment":"This is a complete comment.","section":"Table 1 and Sec. 4.1"}],"minor_comments":[{"comment":"This is a complete comment.","section":"Abstract and Sec. 3.3.1"},{"comment":"This is a complete comment.","section":"Sec. 4.1"},{"comment":"This is a complete comment.","section":"Sec. 4.4.1 and Table 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and addresses a timely problem. The central architectural ideas are plausible, but the empirical evaluation needs to be strengthened before the claims of universal generalization can be accepted. The Texture Function ambiguity is a conceptual issue that should be resolved or bounded in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Alex — quick read of UniTEX. The paper delivers a working two-stage texturing pipeline: LoRA-tuned FLUX for multi-view generation, then a transformer (LTM) that predicts a volumetric Texture Function and bypasses UV maps entirely. The qualitative results are genuinely strong, especially on AI-generated meshes where UV-based inpainting (Paint3D, TexGen) falls apart. The drop-training trick is also nice: 50% token dropout at training, full attention at inference, and they get comparable quality with 44% speedup. That's a practical contribution on its own.\n\nOn the representation side, Texture Functions are close to Oechsle et al.'s Texture Fields, and the paper cites it. The new piece is the truncation, the volumetric supervision to a thin shell, and the transformer architecture. That's an engineering increment, not a new paradigm. The ablation shows TF supervision helps, which is evidence the shell matters.\n\nThe soft spots are real but not fatal. First, Eq. (2) writes the loss as an expectation over surface points, while the text claims dense volumetric supervision. If the implementation samples the volume, the equation is just sloppy notation; if it doesn't, the central claim about volumetric supervision is unsupported. That needs fixing before publication. Second, the TF ground truth is genuinely ill-posed for thin or non-watertight geometry: with a 0.025 truncation, opposing surfaces closer than that create overlapping shells and a discontinuous or ambiguous closest-point map. The paper doesn't analyze this, and the benchmarks lean on solid Objaverse objects, so the failure mode is untested where the method claims its main advantage. That's a missing analysis section, not necessarily a broken method. Third, the quantitative evaluation lacks error bars and the user study is one number with no methodology. The consistent metric gap is reassuring, but the protocol needs detail.\n\nOverall, the engineering holds up and the paper deserves referee time. I'd accept for review but push for a clarification of the training points and a robustness discussion of the TF under thin geometry. Worth citing for the drop-training and the pipeline design, though I'd be careful about citing the TF representation itself without noting the ambiguity.","headline":"Solid two-stage texturing system with a well-motivated UV-free representation; the TF supervision has a real ill-posedness gap the paper doesn't confront, but the empirical results are good enough to warrant a serious review.","tokens_in":14121,"tokens_out":2338,"would_cite":true,"duration_ms":24755,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"UniTEX bypasses UV mapping by predicting textures as continuous 3D functions, outperforming UV-based texturing on both artist-made and AI-generated meshes.","keywords":["texture generation","3D shapes","texture functions","UV mapping","diffusion transformers","LoRA","multi-view consistency","volumetric representation"],"falsifier":"Take a mesh with two closely spaced parallel sheets or a thin open shell, run UniTEX's second stage on it, and inspect the texture at points between the sheets: if the closest-surface-point projection tears or smears the texture, or if the predicted color swaps between the two surfaces at nearby query points, then the claimed topology-agnostic property of Texture Functions fails on realistic geometry.","tokens_in":13081,"feed_emoji":"🎨","tokens_out":2807,"duration_ms":30022,"temperature":0.7,"pith_summary":"This paper tries to establish that UV mapping is a fundamental bottleneck for 3D texture generation, and that replacing it with a continuous volumetric texture representation solves the generalization problem. The authors propose Texture Functions (TFs), which map any 3D point to a texture value based on its closest surface point, making texture representation independent of mesh topology. They build a two-stage pipeline: first, a fine-tuned diffusion transformer generates multi-view images of the mesh; second, a transformer-based Large Texturing Model (LTM) predicts complete texture functions from those images and the geometry. The paper reports that this approach produces more complete, higher-fidelity textures than existing UV-based methods, especially on meshes created by generative AI pipelines.","feed_headline":"Texture functions let 3D texturing skip UV maps entirely","feed_subtitle":"UniTEX predicts textures in 3D space, improving quality on both hand-made and AI-generated meshes.","key_machinery":"The central object is the Texture Function (TF), defined as a mapping $f: \\mathbb{R}^3 \\to \\mathbb{R}^3$ where each point's texture color is inherited from the closest point on the mesh surface. This definition, inspired by unsigned distance functions, extends texture from a surface-restricted signal to a thin volumetric shell, allowing dense 3D supervision. The second key component is the Large Texturing Model (LTM), a transformer that fuses multi-view images, canonical coordinate maps, and geometry into a shared triplane-cube token representation, then decodes queried points into RGB colors through an MLP. A third component is a drop-training strategy for adapting diffusion transformers, which randomly drops input tokens during fine-tuning to reduce memory and speed up training while retaining generation quality.","core_discovery":"The core claim is that UV parameterization introduces a topological ambiguity that prevents texture refinement models from generalizing across the wide variety of mesh topologies found in practice, particularly meshes from generative 3D pipelines. UniTEX avoids this by defining Texture Functions: for any query point in 3D space, the texture value is obtained by orthogonally projecting the point onto the closest surface point of the mesh and reading the color there. This makes the texture representation a continuous, topology-agnostic volumetric field, analogous to how unsigned distance functions represent geometry. A transformer-based Large Texturing Model is trained to regress these texture functions directly from multi-view images and partially textured geometry, using a truncated volumetric supervision loss. The paper demonstrates that this design yields better quantitative and qualitative texture quality than UV-based inpainting methods, particularly in occluded and fragmented regions, and that volumetric texture-function supervision is superior to surface-only supervision.","pith_inferences":["A natural extension beyond this paper is to apply Texture Functions to other surface-bound appearance attributes, such as normal maps, roughness, or metallic PBR channels, which suffer from the same UV-topology issues.","The closest-point projection underlying TFs could be tested as a diagnostic: for meshes with thin shells or intersecting surfaces, measuring the fraction of query points with multiple near-equidistant surface neighbors might predict where texture artifacts appear, something the paper does not analyze.","Because TFs are defined in 3D space rather than on a specific mesh, they could be used to texture implicit surfaces or point clouds directly, potentially unifying texture generation with neural geometry representation.","A testable extension would be to blend TF-based refinement with other completion signals, such as normal-consistency or semantic segmentation, which could improve texture coherence in highly occluded regions beyond what RGB-only supervision achieves."],"forward_implications":["If Texture Functions work as claimed, the dependency on UV unwrapping in texturing pipelines can be removed entirely, eliminating a whole class of topological failure modes.","The volumetric supervision of texture functions suggests that texture can be learned as a continuous field, potentially making texture prediction more robust to imprecise or non-watertight geometry.","The drop-training strategy indicates that large diffusion transformers can be adapted to multi-view conditioning tasks with significantly lower computational cost, which could accelerate other 3D generation tasks.","The reported gains on generative meshes imply that texturing can keep pace with fast 3D shape generation pipelines, a necessary step for automated asset creation.","The separation into a view-generation stage and a 3D refinement stage offers a modular recipe: improvements in either 2D diffusion priors or 3D texture representations can be integrated independently."],"supporting_citations":[{"why":"Paint3D is the primary UV-based texturing baseline whose inpainting failure on fragmented UV layouts motivates the need for UV-free texture functions.","marker":"[40]"},{"why":"TexGEN is a recent native 3D and UV-based texturing method used as a second-stage comparison to show the limitations of UV-space refinement.","marker":"[39]"},{"why":"Hunyuan3D 2.0 is a state-of-the-art multi-view texturing baseline that the paper compares against on both artist-created and generative meshes.","marker":"[47]"},{"why":"Meta 3D TextureGen represents the geometry-aware multi-view synthesis approach that the paper positions against in the related work and evaluation.","marker":"[1]"},{"why":"Texture Fields introduced implicit color fields on 3D surfaces, the conceptual precursor to the proposed volumetric texture functions.","marker":"[27]"},{"why":"Craftsman is the 3D mesh generation pipeline used to create the generative test meshes, establishing the distribution that UV-based methods struggle with.","marker":"[19]"},{"why":"CRM supplies the canonical coordinate map and triplane encoding technique used to initialize the triplane-cube representation in the Large Texturing Model.","marker":"[36]"},{"why":"Shape2VecSet provides the geometry tokenization method used to query the triplane-cube with colored point cloud features.","marker":"[41]"},{"why":"MVDiffusion++ demonstrates that training on fewer views can generalize to more views at inference, a key inspiration for the drop-training strategy.","marker":"[34]"},{"why":"LongLoRA contributes the idea of truncating attention during training while restoring full attention at inference, directly motivating the token dropout approach.","marker":"[6]"}],"fun_headline_variants":["UniTEX paints 3D shapes without UV maps","Topology-agnostic textures via 3D texture functions","Skip UV mapping: texturing in 3D space","Predict 3D textures directly from images and meshes","Continuous texture fields replace UV unwrapping"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Texture Function representation assumes that every 3D query point has a single closest surface point, but for thin structures, overlapping surface layers, or non-watertight meshes this projection can be ambiguous or discontinuous, and the paper does not analyze this failure mode.","fun_headline_variants_meta":{"raw":{"variants":["UniTEX paints 3D shapes without UV maps","Topology-agnostic textures via 3D texture functions","Skip UV mapping: texturing in 3D space","Predict 3D textures directly from images and meshes","Continuous texture fields replace UV unwrapping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000631,"raw_usage":{"total_tokens":2928,"prompt_tokens":969,"completion_tokens":1959,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":1880}},"tokens_in":585,"tokens_out":1959,"duration_ms":14723,"temperature":1.0,"reasoning_tokens":1880,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:48:37.193921+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a mesh with two closely spaced parallel sheets or a thin open shell, run UniTEX's second stage on it, and inspect the texture at points between the sheets: if the closest-surface-point projection tears or smears the texture, or if the predicted color swaps between the two surfaces at nearby query points, then the claimed topology-agnostic property of Texture Functions fails on realistic geometry.","supporting_citations":[{"cited_title":"Paint3d: Paint anything 3d with lighting-less texture diffusion models","cited_arxiv_id":null,"evidence_quote":"Paint3D is the primary UV-based texturing baseline whose inpainting failure on fragmented UV layouts motivates the need for UV-free texture functions."},{"cited_title":"Texgen: a generative diffusion model for mesh textures","cited_arxiv_id":null,"evidence_quote":"TexGEN is a recent native 3D and UV-based texturing method used as a second-stage comparison to show the limitations of UV-space refinement."},{"cited_title":"Texture fields: Learning tex- ture representations in function space","cited_arxiv_id":null,"evidence_quote":"Texture Fields introduced implicit color fields on 3D surfaces, the conceptual precursor to the proposed volumetric texture functions."},{"cited_title":"3dshape2vecset: A 3d shape representation for neu- ral fields and generative diffusion models.ACM Transactions on Graphics (TOG), 42(4):1–16, 2023","cited_arxiv_id":null,"evidence_quote":"Shape2VecSet provides the geometry tokenization method used to query the triplane-cube with colored point cloud features."},{"cited_title":"Mvdiffusion++: A dense high- resolution multi-view diffusion model for single or sparse- view 3d object reconstruction","cited_arxiv_id":null,"evidence_quote":"MVDiffusion++ demonstrates that training on fewer views can generalize to more views at inference, a key inspiration for the drop-training strategy."}],"review_version":1}