{"id":"4234d626-43cf-4b7f-8fdd-cb4209dcafcc","arxiv_id":"2509.05131","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A DINOv2-conditioned transformer decodes a triplane texture field on a given mesh, supervised by depth backprojection, producing UV-free textures in 0.2 s per shape.","lead":"A single image plus a 3D mesh go into a transformer that paints the mesh with colors in one 0.2 second pass, without UV maps. The paper claims this is much faster and more faithful than diffusion-based texturing tools, but its evidence needs stronger comparisons.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fixed 55-view camera rig in training/evaluation means image-to-mesh alignment is never tested; arbitrary-view or real-photo use of the central claim is unsupported.","rationale":"I read the paper in good faith and find the architecture to be a reasonable and well-motivated adaptation of LRM to UV-free texture fields. The speed advantage is plausible and the ablations are informative. The reader's weakest assumption—unverified image-mesh alignment—is the same concern I identify, and I agree with the conditional verdict. However, I want to sharpen it: the alignment problem is not only about inaccurate depth/pose or misaligned arbitrary sources. The paper's own protocol may already avoid the difficult part by training and testing with a fixed set of 55 camera viewpoints (Section 4.1). Since the model receives no pose or geometry conditioning, it can only learn correspondences from the distribution of those views. The '10 novel views' in Table 1 are novel objects, but there is no evidence they are novel viewpoints. If they come from the same 55-view rig, the evaluation systematically underestimates the risk of viewpoint overfitting. This does not make the method fraudulent or the architecture unsound; it makes the central claim narrower than stated. I also note the baselines are optimization-based and not all are natively image-guided, and TEXTure's training steps were reduced, which weakens the quantitative comparison; but that is secondary to the generalization gap. If the random-viewpoint test I propose passes, the central claim is substantially supported; if it fails, the claim should be re-scoped to aligned rendered views or extended with pose/geometry conditioning. Because the reader already arrived at conditional acceptance, my stress test does not change the verdict; it provides a concrete experiment to resolve the uncertainty.","tokens_in":11213,"tokens_out":7429,"duration_ms":86565,"concrete_test":"Render each of the 100 GSO test objects from 20 random camera viewpoints sampled continuously in azimuth/elevation (not among the 55 precomputed views), with the same focal length distribution, and use each render as the conditioning image while keeping the mesh fixed. Compute PSNR/LPIPS/CLIP on the same held-out novel views used in Table 1. If mean PSNR drops by more than ~3 dB from 27.65 or LPIPS exceeds ~0.12, the model has overfit to the fixed training camera rig and does not perform general single-image texture mapping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 trains solely by backprojecting precomputed depth maps with known intrinsics K and camera-to-world T_cam, and Section 3.1 explicitly drops geometric conditioning. The transformer thus receives no camera pose, no geometry, and no explicit correspondence signal; any ability to align an arbitrary conditioning image to object-space triplane coordinates must be learned implicitly from the training-view distribution. Section 4.1 precomputes RGB/depth from exactly 55 viewpoints for both Objaverse and GSO, and the GSO evaluation (Table 1) uses 10 'novel views' per object without stating that these views or the conditioning image fall outside the same 55-view rig. Therefore the central claim—that a single forward pass maps a single image and a mesh to a faithful texture—is only demonstrated in the aligned-rendered-view setting. If the conditioning image comes from a different camera pose, or from a real photo with unknown intrinsics/extrinsics, the implicit correspondence may break; no experiment or analysis rules this out. This is load-bearing because the advertised practical applicability of 'single-image texture reconstruction' depends on unposed image-to-mesh alignment, which the paper does not test.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a feed-forward transformer-based method for texturing a given 3D mesh from a single image. A DINOv2 encoder extracts image features; a transformer decoder with learned positional embeddings and cross-attention produces a 32×32 triplane, upsampled to 64×64; an MLP decodes triplane features to RGB at arbitrary 3D points. Training supervises the texture field by backprojecting precomputed depth maps from 55 viewpoints and comparing predicted and ground-truth images with L2 and LPIPS losses. Inference is a single forward pass, reported at ~0.2 s per mesh with no UV mapping. Experiments compare to TEXTure, Paint3D, and EASI-Tex on GSO and Objaverse, including quantitative metrics, ablations on model size/conditioning/losses, and a user study.","tokens_in":11542,"tokens_out":7133,"duration_ms":77360,"significance":"If the claims hold, the method is a useful step toward practical, UV-free, fast texture generation from images, and the integration of triplanes with transformer conditioning is a reasonable architectural choice. Strengths include the 0.2 s feed-forward inference, cross-dataset evaluation on GSO, ablations, and a user study. However, the empirical support is weakened by an under-trained TEXTure baseline, missing feed-forward comparisons, and an unexamined image-mesh alignment assumption; these need to be addressed before the headline claims are fully supported.","major_comments":[{"comment":"The text states that baselines are 'running each baseline to convergence', but the TEXTure baseline uses max train steps reduced from 10,000 to 1,000. This is 10% of the default and likely produces an undertrained model. The reported margins over TEXTure (CLIP 90.09 vs 80.24, LPIPS 0.075 vs 0.236, PSNR 27.65 vs 13.31) may therefore substantially overstate the improvement. Please provide a convergence-controlled comparison, or report TEXTure at its default step count and justify why 1,000 steps is sufficient.","section":"§4.1, Table 1"},{"comment":"Section 2.4 surveys feed-forward methods (LRM, Instant3D, GRM), but the experimental comparison includes only iterative optimization baselines (TEXTure, Paint3D, EASI-Tex). Given the paper's central claims are feed-forward speed and scalability, the absence of any feed-forward baseline leaves the 'state-of-the-art' claim unsupported. Please add at least one feed-forward method applicable to known-mesh texturing, or explicitly state why no such baseline is suitable.","section":"§4.1, Table 1"},{"comment":"The model receives no camera pose or geometric conditioning (Section 3.1 explicitly drops geometric conditioning), yet training and evaluation data are rendered from a fixed 55-view rig. For GSO, the conditioning image is a rendered view of the same mesh, and no experiment uses a real photograph or an arbitrary viewpoint with unknown pose. The image-to-object correspondence must therefore be learned implicitly from the training-view distribution, and the paper does not test whether it transfers outside that distribution. This is load-bearing for the advertised 'single-image texture reconstruction' claim. Please run held-out-viewpoint experiments, or at minimum explicitly scope the claim to aligned rendered views.","section":"§3.1, §3.4, §4.1"},{"comment":"The training protocol is underspecified: the text says 'we process 4 random views from a set of 55 precomputed depth maps' and the loss sums over V views, but it never states which view is used as the conditioning image and whether that view is included in the four supervised views. The architecture (Section 3.3) and Figure 2 show a single conditioning image. If the conditioning view is always one of the supervised views, the model is trained to reconstruct an observed view plus nearby views, not to texture from a truly arbitrary single image. Please clarify the protocol and, ideally, exclude the conditioning view from the supervised views in evaluation.","section":"§3.4"}],"minor_comments":[{"comment":"DINOv2 outputs patch tokens; the paper writes z∈R^768 as a single vector. Please clarify whether this is a pooled/CLS-like feature and how it is computed.","section":"§3.3"},{"comment":"The inference loop says 'for each query point p∈M' but does not specify how query points are sampled or at what density. This affects runtime and texture quality; please specify the sampling strategy.","section":"Algorithm 1"},{"comment":"For the GSO evaluation, please state explicitly whether the 10 novel views and the conditioning image are disjoint from the 55-view rig used in training/precomputation. This is important for interpreting the single-image generalization claim.","section":"§4.1"},{"comment":"The text says 'Our approach outperforms both baselines' but Table 1 lists three baselines (TEXTure, Paint3D, EASI-Tex). Please correct this inconsistency.","section":"§4.2"},{"comment":"The failure cases are presented only qualitatively. A short quantitative analysis (e.g., resolution or frequency-band metrics) would make the limitation more precise. Also, the caption has a typo: 'Failure casesWhile'.","section":"§4.4 and Figure 7"},{"comment":"The geometric conditioning baseline called 'Latent' [34] relies on a reference that is not described in the paper and is a self-citation. Please specify what the latent representation is and why it is a reasonable SDF feature source.","section":"§4.4"},{"comment":"References [45] and [46] appear to be the same work (Zhang et al., 'The unreasonable effectiveness of deep features as a perceptual metric'). Please merge the duplicate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the empirical comparison: the TEXTure baseline under-training and the lack of a feed-forward baseline could substantially change the reported superiority. The alignment/viewpoint issue is also important for scoping the claim, but it may be addressable with additional experiments or an explicit limitation statement. The method itself is plausible and the architectural choices are reasonable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a competent feed-forward texture-field paper with a real speed advantage, but the evaluation rig means the headline claim—single image + mesh → faithful texture—is only tested in the aligned-rendered-view setting. The stress-test note is correct: arbitrary image-to-mesh alignment is never exercised.\n\nWhat's new: the combination of DINOv2 conditioning, a triplane decoder, and depth-backprojection supervision for texture-only prediction on known meshes is not in the cited prior work. The paper explicitly concedes the components are established, and the ablation showing geometric conditioning doesn't help is a genuinely useful negative result. The model-size study is honest: BASE saturates, LARGE adds little. Speed (~0.2s on A10) is consistent with the architecture. Failure cases are shown, which is more than many papers do.\n\nSoft spots: the main one is the fixed 55-view rig. Training backprojects depth with known K and T_cam, and the model receives no pose or geometry, so any ability to align an arbitrary conditioning image must be learned implicitly from the training-view distribution. No experiment tests a different pose or a real photo. That is a load-bearing gap for the practical claim. Second, the baseline comparison is not fair: TEXTure runs at 1,000 of its default 10,000 training steps, contradicting the 'to convergence' statement. Also, no feed-forward method (LRM, GRM, etc.) is compared, which is the most relevant category. Third, no code or uncertainty estimates are provided, which limits falsifiability. The mild test-train overlap with GSO is not a big deal to me, but the same render pipeline is worth noting.\n\nProportion: the architecture is plausible and the speed claim holds. The quantitative 'outperforms state-of-the-art' claim is weakened by the baseline issues, but the core idea is sound.\n\nWho this is for: researchers working on feed-forward 3D texture generation or LRM-style reconstruction. It deserves a serious referee. My recommendation: send to peer review, but require a revision that adds arbitrary-view or real-photo conditioning tests and fair baselines. I'd bring it to a reading group to discuss evaluation protocols.","headline":"Solid LRM-style texture-field adaptation with a real speed win, but the fixed 55-view evaluation rig means arbitrary image-to-mesh alignment is never tested, so the headline claim is narrower than advertised.","tokens_in":12002,"tokens_out":2563,"would_cite":true,"duration_ms":29667,"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":"A transformer predicts a full triplane texture field for any mesh from a single image in a single forward pass, eliminating UV maps and optimization.","keywords":["texture field","triplane","transformer","single-image 3D reconstruction","mesh texturing","depth backprojection","Objaverse","feed-forward"],"falsifier":"Run the trained model on real photographs of objects paired with meshes that are not rendered from those same views, or where depth comes from a monocular estimator instead of ground-truth precomputed depth maps; if PSNR/LPIPS on this test collapses compared to the paper's GSO numbers, the method's generality depends on the image being a known rendered view of the same mesh.","tokens_in":11153,"feed_emoji":"🎨","tokens_out":3644,"duration_ms":41320,"temperature":0.7,"pith_summary":"This paper claims that texturing a known 3D mesh from a single reference image can be solved as a feed-forward prediction problem rather than a per-object optimization. The method encodes the image with a pretrained vision transformer, decodes a triplane feature field through a transformer with cross-attention, and supervises the result by backprojecting depth maps into 3D point clouds with ground-truth colors. The central result is that this UV-free, differentiable-rendering-free pipeline reconstructs textures that are more faithful to the input image and more perceptually similar to ground truth than leading diffusion-based optimization baselines, while running in about 0.2 seconds per mesh instead of tens of minutes. If correct, texture generation becomes fast enough for large-scale asset pipelines and robust to meshes that lack clean UV parametrization.","feed_headline":"A single image textures a 3D mesh in 0.2 seconds","feed_subtitle":"Transformer-based triplane field beats diffusion baselines on fidelity and runs orders of magnitude faster.","key_machinery":"The key mechanism is a triplane texture field decoded by a transformer: learned positional embeddings are passed through a transformer decoder that cross-attends to image features from a self-supervised vision encoder, then reshaped into three orthogonal feature planes (XY, XZ, YZ). Any 3D point on the mesh is projected onto the three planes, bilinearly sampled, concatenated, and decoded to RGB by a small MLP. This triplane field is supervised entirely through depth-map backprojection, which converts each pixel's depth and camera pose into a 3D query point with a ground-truth color, avoiding the need for UV maps or differentiable rasterization.","core_discovery":"The paper establishes that a continuous texture field over known geometry can be learned end-to-end from a single conditioning image without any UV mapping or differentiable rendering. The model, a transformer-based triplane decoder, maps image features to a 3D texture field T(p, I) that answers arbitrary 3D point queries with RGB colors. Supervision comes from depth-map backprojection: precomputed depth maps and camera poses convert each pixel into a 3D point carrying its ground-truth color, and the predicted field is penalized with a pixel-wise loss plus LPIPS. On the Google Scanned Objects benchmark the method reports CLIP-Score 90.09, LPIPS 0.075, and PSNR 27.65, substantially outperform","pith_inferences":["A natural extension, hinted at in the paper's future-work section, is to chain this fast feed-forward texture field as an initialization for a diffusion-based refinement stage, which could recover the high-frequency detail that the current model misses without sacrificing speed.","The depth-backprojection supervision could be made fully general by replacing precomputed ground-truth depth with monocular depth estimates during training or inference, which would let the method handle arbitrary real photographs rather than rendered views of the same mesh.","The same triplane field formulation could support multi-image conditioning by concatenating several cross-attention streams, potentially improving fidelity for complex or occluded objects beyond a single view.","Because the model learns a class-agnostic mapping from image to texture field, it may support texture transfer between different geometries that share semantic appearance, a capability the paper does not explore."],"forward_implications":["Texture generation drops from minutes or hours of per-object optimization to roughly 0.2 seconds per mesh, making it practical to texture thousands of assets in a batch.","Because the method never requires a UV map, it can texture meshes produced by generative geometry models or real-world scans whose topology lacks a clean 2D parametrization.","The texture field is queried pointwise, so the same trained model works at any mesh resolution by sampling more points during inference.","The reported margins over diffusion-based baselines on CLIP, LPIPS, and PSNR imply that, for image-conditioned texture reconstruction, a direct feed-forward mapping can beat iterative optimization while being far faster.","The saturation point in model capacity (BASE over LARGE) suggests that the remaining bottleneck in texture quality is representation capacity or supervision, not just transformer size."],"supporting_citations":[{"why":"Establishes the transformer-plus-triplane architecture for single-image 3D reconstruction that this method adapts by separating texture from geometry.","marker":"[13]"},{"why":"Introduces the texture-field concept of learning a continuous 3D color function, which this work scales beyond single-category data.","marker":"[24]"},{"why":"Introduces the triplane representation used as the paper's intermediate 3D feature field.","marker":"[3]"},{"why":"Provides the self-supervised image encoder that produces the visual conditioning for the transformer decoder.","marker":"[25]"},{"why":"Supplies the LPIPS perceptual loss that the method uses to balance fidelity and perceptual quality in supervision.","marker":"[46]"},{"why":"The Objaverse dataset is the large-scale training source with diverse textured 3D assets.","marker":"[8]"},{"why":"The Google Scanned Objects benchmark is the out-of-domain real-world scanned meshes used for evaluation.","marker":"[10]"},{"why":"Baseline diffusion-based optimization texturing method that requires UV maps and fine-tuning, used as a comparison.","marker":"[31]"},{"why":"Baseline multi-view diffusion and UV inpainting method with test-time refinement, used as a comparison.","marker":"[43]"},{"why":"Baseline edge-aware mesh texturing from a single image, used as a comparison and in the user study.","marker":"[26]"}],"fun_headline_variants":["Texture a 3D mesh from one image in 0.2s","0.2-sec image-to-3D texture without UV maps","Transformer paints 3D meshes from a single image","Single-image 3D texture mapping in 0.2 seconds","No UV, no rendering: image to 3D texture in 0.2s"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole supervision signal assumes every conditioning image is a rendered view of the exact mesh being textured, with depth and camera pose already known; if image and mesh are not aligned this way, the learned correspondences break.","fun_headline_variants_meta":{"raw":{"variants":["Texture a 3D mesh from one image in 0.2s","0.2-sec image-to-3D texture without UV maps","Transformer paints 3D meshes from a single image","Single-image 3D texture mapping in 0.2 seconds","No UV, no rendering: image to 3D texture in 0.2s"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00032,"raw_usage":{"total_tokens":1614,"prompt_tokens":694,"completion_tokens":920,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":833}},"tokens_in":438,"tokens_out":920,"duration_ms":8269,"temperature":1.0,"reasoning_tokens":833,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:33:41.814561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on real photographs of objects paired with meshes that are not rendered from those same views, or where depth comes from a monocular estimator instead of ground-truth precomputed depth maps; if PSNR/LPIPS on this test collapses compared to the paper's GSO numbers, the method's generality depends on the image being a known rendered view of the same mesh.","supporting_citations":[{"cited_title":"Lrm: Large reconstruction model for single image to 3d, 2024","cited_arxiv_id":null,"evidence_quote":"Establishes the transformer-plus-triplane architecture for single-image 3D reconstruction that this method adapts by separating texture from geometry."},{"cited_title":"Texture fields: Learning tex- ture representations in function space, 2019","cited_arxiv_id":null,"evidence_quote":"Introduces the texture-field concept of learning a continuous 3D color function, which this work scales beyond single-category data."},{"cited_title":"Efficient geometry-aware 3d gen- erative adversarial networks.” arxiv, 2021","cited_arxiv_id":null,"evidence_quote":"Introduces the triplane representation used as the paper's intermediate 3D feature field."},{"cited_title":"Dinov2: Learning robust visual features with- out supervision, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the self-supervised image encoder that produces the visual conditioning for the transformer decoder."},{"cited_title":"Google scanned objects: A high- quality dataset of 3d scanned household items","cited_arxiv_id":null,"evidence_quote":"The Google Scanned Objects benchmark is the out-of-domain real-world scanned meshes used for evaluation."},{"cited_title":"Texture: Text-guided texturing of 3d shapes, 2023","cited_arxiv_id":null,"evidence_quote":"Baseline diffusion-based optimization texturing method that requires UV maps and fine-tuning, used as a comparison."},{"cited_title":"Paint3d: Paint anything 3d with lighting-less texture diffusion models, 2023","cited_arxiv_id":null,"evidence_quote":"Baseline multi-view diffusion and UV inpainting method with test-time refinement, used as a comparison."},{"cited_title":"Easi-tex: Edge-aware mesh texturing from sin- gle image.ACM Transactions on Graphics (TOG), 43(4): 1–11, 2024","cited_arxiv_id":null,"evidence_quote":"Baseline edge-aware mesh texturing from a single image, used as a comparison and in the user study."}],"review_version":1}