{"id":"297bfa56-d139-420e-9d81-2dc7a1556aa9","arxiv_id":"2506.10612","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A diffusion-based texturing method that uses DDIM resampling, per-object ControlNet fine-tuning with a preservation loss, and adaptive camera placement to reduce view-to-view texture drift.","lead":"TexTailor generates textures for 3D objects from text by repeatedly blending previously painted views inside a diffusion model, then fine-tuning the model on a handful of those blended views so later angles match. The result is a texturing pipeline that keeps colors and patterns more consistent across viewpoints than earlier methods on standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative support for the central claim rests on an LPIPS metric that the paper itself (A.4) admits does not capture view consistency; Table 1 and Table 3 also disagree by ~4x. A valid consistency metric is needed before 'outperforms SOTA' can be assessed.","rationale":"The reader's verdict of CONDITIONAL is reasonable, but the weakest assumption they identified—fine-tuning on only five images—is not the most load-bearing issue for the central claim. That assumption affects worst-case behavior and is explicitly acknowledged in A.4, but even in the nominal case the quantitative evidence for 'outperforms SOTA in view consistency' is invalidated by the paper's own admission that LPIPS is not a valid consistency measure. Additionally, the unexplained factor-of-four difference between Table 1 and Table 3 suggests the evaluation protocol is unstable, so the numbers cannot be trusted without clarification. A stronger metric and a reproducible protocol are prerequisites for assessing the headline claim; until then the paper is unverdictable with respect to its central contribution. I therefore recommend UNVERDICTED rather than CONDITIONAL or REJECT, because the methods may still work and the qualitative results are suggestive, but the current evidence does not establish the claimed superiority. The proposed concrete test—re-evaluating with a geometry-aware correspondence metric and reconciling Table 3 with Table 1—would resolve this concern directly.","tokens_in":13518,"tokens_out":4911,"duration_ms":60730,"concrete_test":"Recompute Table 1 using a geometry-aware view-consistency metric: render each object from the same camera set, establish dense correspondences across views (e.g., via the known mesh UV mapping or optical flow), and measure consistency only on corresponding surface points (e.g., mean per-pixel RGB or Deep feature distance at matched coordinates). Also recompute Table 3 under exactly the same protocol as Table 1 to resolve the ~4x LPIPS discrepancy. If TexTailor no longer ranks first, or if the Table 1/Table 3 numbers cannot be reproduced, the quantitative claim of state-of-the-art view consistency is unsupported.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"TexTailor's central claim—'outperforms state-of-the-art methods in synthesizing view-consistent textures'—is supported primarily by Table 1, which reports average LPIPS and FID on an Objaverse subset. But Appendix A.4 explicitly states: 'LPIPS does not adequately capture consistency across multiple viewpoints due to spatial misalignments in overlapping sections between adjacent views.' If the paper's own chosen metric is acknowledged to be inadequate for measuring the claimed property, then Table 1 cannot establish the central claim. The issue is compounded by a large, unexplained quantitative inconsistency: Table 1 reports LPIPS values around 38–54 for all methods, while Table 3—using the same metric on the same method—reports values around 8.99. FID values likewise jump from ~30–45 in Table 1 to ~57–59 in Table 3. No change in protocol (e.g., different subset, render resolution, or normalization) is described that would reconcile these numbers. This makes the quantitative evidence internally unstable and not reproducible. The five-image fine-tuning assumption flagged in the reader's verdict is a real failure mode, but it concerns robustness to training-image quality; the metric problem directly undermines the proof that the method beats prior work at all. Without a geometry-aware or otherwise validated consistency metric, the headline outperformance claim is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents TexTailor, a method for text-to-texture synthesis on 3D meshes. It proposes three contributions: (1) a resampling scheme applied within the DDIM non-Markovian process to repeatedly integrate previously synthesized textures during inpainting-based texturing; (2) per-object fine-tuning of a ControlNet depth-aware diffusion model on a small set of resampled images, regularized by a 'performance preservation loss' to mitigate catastrophic forgetting; and (3) adaptive viewpoint refinement that inserts intermediate camera positions based on the ratio of kept to new texture regions. The method is evaluated on a subset of Objaverse and qualitatively on ShapeNet cars, reporting lower LPIPS and FID than prior methods, with ablations showing that each component contributes to the final result.","tokens_in":13820,"tokens_out":8967,"duration_ms":92927,"significance":"If the quantitative claims are substantiated, TexTailor offers a practical and interesting recipe for improving view consistency in text-to-texture generation without requiring external 3D datasets: the model is fine-tuned on its own resampled outputs while a distillation-style loss keeps it close to the pretrained model. The promise of released source code, the per-component ablations, and the adaptive viewpoint mechanism are clear strengths. However, the current evidence is weakened by an underspecified core equation, internally inconsistent quantitative tables, and a metric that the authors themselves acknowledge does not measure the claimed property. With a valid consistency metric and a cleaned-up presentation, the contribution could be solid and useful to the community.","major_comments":[{"comment":"Eq. (7) uses the quantity \\tilde z^r_0, which is never defined. The preceding text defines \\tilde z^0_{t-1} (as the result of Eq. 5) and \\tilde z^r_t (via Eq. 6), but the DDIM denoising formula requires an estimate of the clean latent, presumably \\tilde z^r_0 = (\\tilde z^r_t - \\sqrt{1-\\bar\\alpha_t}\\,\\epsilon_\\phi(\\tilde z^r_t, t, c))/\\sqrt{\\bar\\alpha_t}. Without this definition, the resampling recursion is not reproducible. Please define \\tilde z^r_0 explicitly and verify the subscripts throughout Eqs. (6)-(8).","section":"Sec. 3.1, Eq. (7)"},{"comment":"The quantitative results are internally inconsistent: Table 1 reports LPIPS values around 38-54 and FID around 30-45, while Table 3, using the same metric names, reports LPIPS around 8.99 and FID around 57-59 for the same method. No difference in protocol (e.g., different mesh subset, render resolution, camera sampling, or LPIPS implementation) is described to reconcile these numbers. As written, the discrepancy undermines the credibility of the comparisons in Table 1. Please clarify the protocols and, if the numbers come from different settings, explain precisely why the scales differ by a factor of roughly four.","section":"Sec. 4.1 and A.6, Tables 1 and 3"},{"comment":"The central claim is that TexTailor improves view consistency, but the only quantitative metric used to support this claim, LPIPS between rendered views, is explicitly stated in A.4 to 'not adequately capture consistency across multiple viewpoints due to spatial misalignments in overlapping sections between adjacent views.' Thus Table 1 does not measure the property it is used to assert. Please add a directly consistency-aware metric, such as warped-view agreement with depth, texture-map alignment error, or a user study, and report it for TexTailor and all baselines. Without such a metric, the headline outperformance claim is unverified.","section":"Sec. 4.1, Table 1 and A.4"},{"comment":"The description of the LPIPS aggregation says: 'compute the average LPIPS values for all pairs of images in the 3D scene and sum the averages across all evaluated categories.' Summing rather than averaging over categories makes the score dependent on the number of categories and objects, and it is unclear whether the numbers in Tables 1 and 2 are sums or means. This also prevents comparison across datasets of different sizes. Please clarify the aggregation and, ideally, report the mean and standard error over objects.","section":"Sec. 4.1, evaluation metric aggregation"},{"comment":"The method's success depends heavily on the five fine-tuning images, and A.4 concedes that 'suboptimal training images can sometimes degrade texture quality rather than enhance it.' However, no quantitative sensitivity analysis is provided: there is no report of variance across different random seeds, different choices of the five viewpoints, or different objects where the training images are poor. Given that this fine-tuning step is a core component of the pipeline, please add such an analysis or explicitly characterize the conditions under which the method is expected to fail.","section":"Sec. 4.1, Sec. 4.2, and A.4"}],"minor_comments":[{"comment":"In Eq. (3), z^{known}_{t-1} is defined using \\bar\\alpha_t, but the standard forward-process posterior for timestep t-1 should use \\bar\\alpha_{t-1}; please check the subscript.","section":"Sec. 2.2, Eq. (3)"},{"comment":"The sentence 'geometry optimization process (Shen et al., 2021; ?; Nash et al., 2020)' contains a stray '?' placeholder; please remove or replace it with the intended citation.","section":"Sec. 1, first paragraph"},{"comment":"The phrase 'an performance preservation loss' appears in the abstract, Sec. 1, Sec. 3.2, and elsewhere; it should be 'a performance preservation loss.' Similarly, 'we finetunes' in Sec. 1 should be 'we fine-tune.'","section":"Throughout"},{"comment":"The text says 'randomly select 100 meshes from the Objaverse dataset, ensuring no overlapping categories among the 400 meshes,' but Sec. 4.1 states the Objaverse subset contains 410 meshes. Please correct this number or explain the discrepancy.","section":"Appendix A.6"},{"comment":"The sentence 'visual differences are harder to discern in the qualitatve comparison sections (Sec. 5 and Sec. 8)' contains a typo ('qualitatve') and references sections numbered 5 and 8, which do not exist in the paper's numbering; please correct the cross-references.","section":"Appendix A.5"}],"recommendation":"major_revision","confidential_remarks":"The paper is an ICLR 2025 submission posted on arXiv. My main concern is the quantitative evaluation: the LPIPS inconsistency between Tables 1 and 3 and the acknowledged inadequacy of LPIPS for the central claim would need to be addressed before the paper can be accepted. The core ideas are plausible and the ablations are informative, so a major revision with a valid consistency metric and clarified equations seems appropriate. I do not see evidence of deliberate misreporting, but the current numerical discrepancies must be resolved by the authors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The method is a sensible combination of known tools, and the qualitative story is believable. But the numbers in the main table cannot carry the claim: the paper's own limitation section says LPIPS doesn't measure view consistency, and the same metric appears to produce values around 38–54 in Table 1 and around 9 in Table 3 with no explanation. That's a 4x gap on the headline metric, and FID moves the other way. The 'outperforms SOTA' sentence is not supported by the evidence as written.\n\nWhat's genuinely new: applying RePaint-style resampling inside DDIM for texturing, per-object ControlNet fine-tuning with a preservation loss to fight forgetting, and adaptive camera placement based on keep-region ratio. The ablations show each component helps, and the qualitative figures, especially the muffin and basket, make a plausible case that the method reduces gradual drift. The authors also earn credit for an honest limitations section; they flag the five-image fragility and the LPIPS issue themselves.\n\nSoft spots, in order. First, the metric problem is real and load-bearing. If the main metric is admitted not to measure the claimed property, Table 1 cannot establish the result. The Table 1 vs Table 3 discrepancy also suggests some unstated protocol change, which undermines reproducibility. Second, Eq. 7 uses z-tilde-r-0 without defining it. You can guess it is the predicted clean latent from a previous resampling step, but guessing is not what a reader should have to do for the central equation. Third, fine-tuning on five self-generated images is a genuine robustness risk; the paper acknowledges it can degrade quality. For some objects it clearly works, but on a harder mesh it could be a coin flip. Fourth, the comparison is not controlled: per-object fine-tuned model against frozen baselines, no error bars, and only one prompt template.\n\nWho is this for? People working on text-to-texture who want a concrete recipe to reduce drift. The recipe is worth trying, and the adaptive viewpoint idea could be reused independently. But the paper should not be cited for the quantitative superiority claim until the metric issue is resolved.\n\nIf this crossed my desk, I would send it to reviewers. The core approach is sound enough to warrant a serious look, and the fixes are specific: report a geometry-aware consistency metric, reconcile the tables, and define the resampling notation.","headline":"Plausible method, honest limitations, but the main metric is admitted not to measure view consistency and the tables disagree by 4x; worth reviewing, not worth citing for the numbers.","tokens_in":14351,"tokens_out":3014,"would_cite":true,"duration_ms":33926,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TexTailor claims view-consistent 3D textures come from resampling during denoising, few-shot model fine-tuning, and adaptive camera placement.","keywords":["text-to-texture synthesis","view consistency","diffusion models","DDIM resampling","fine-tuning with preservation loss","adaptive viewpoint refinement","3D mesh texturing","catastrophic forgetting"],"falsifier":"Render a mesh with strong front/back asymmetry using a deliberately corrupted first-view training image, then compare the texture at the opposite viewpoint to the uncorrupted run: if the opposite view does not visibly drift or degrade, the fine-tuning step is more robust than the paper's stated five-image dependence.","tokens_in":13298,"feed_emoji":"🎨","tokens_out":11686,"duration_ms":115742,"temperature":0.7,"pith_summary":"TexTailor is a text-to-texture method for meshes that tries to stop the gradual drift in color and pattern that appears as a texture is painted view by view. Its central claim is that the drift comes from under-integrating previously painted texture during denoising and from an autoregressive synthesis order, and that both can be fixed with three moves: a resampling loop inside DDIM that repeatedly merges known and unknown regions, per-object fine-tuning of a depth-conditioned diffusion model on five resampled images with a preservation loss, and adaptive addition of camera positions when too little previously painted texture is visible. On a subset of Objaverse and on ShapeNet cars, the paper reports lower LPIPS and FID than existing text-driven texture methods, meaning textures that look more alike across views. A sympathetic reader would take the contribution to be a practical recipe for adapting a pretrained 2D diffusion model to a single object's texture while keeping the model's original generation ability.","feed_headline":"Resampling stops texture drift across 3D views","feed_subtitle":"Repeated merging of painted and unpainted regions, plus five-view training, keeps textures consistent across angles.","key_machinery":"The engine of the method is resampling in a DDIM non-Markovian process. At each denoising step, instead of merging the known (already painted) and unknown (to-be-painted) regions once, TexTailor adds noise to the merged latent, denoises it, and merges again, repeating $R$ times; this integrates previously synthesized texture into the new region with only 30 steps per view. On top of that, a depth-conditioned diffusion model (ControlNet) is fine-tuned on five resampled images near the first viewpoint, with a preservation loss $\\mathcal{L}_{\\mathrm{pre}}$ that penalizes deviation from the frozen pretrained noise predictor to prevent catastrophic forgetting. The third mechanism is an adaptive viewpoint rule: when the ratio $p$ of \"keep\" pixels to (keep + new) pixels falls below a threshold $\\beta$, an interpolated camera position is inserted, so the next view always has enough previously painted texture to condition on.","core_discovery":"On the paper's own terms, the discovery is that the gradual texture shift in multi-view texturing is not inevitable: it can be suppressed by repeating the known/unknown merge several times per diffusion timestep inside a non-Markovian sampler, then teaching the depth-aware model the object's appearance from a handful of resampled views. The fine-tuning step is the load-bearing addition because resampling alone still lets properties drift at viewpoints far from the first camera. To keep that fine-tuning from destroying the pretrained model's fidelity, the paper adds a preservation loss that pulls the fine-tuned noise prediction toward the original pretrained noise prediction. The final claim is that the measured result, lower average LPIPS and FID than the listed baselines on an Objaverse subset plus qualitative gains on ShapeNet cars, supports the whole combination.","pith_inferences":["Not claimed by the paper: the same preservation-loss term could serve as a general regularization for any few-shot fine-tuning of conditioned diffusion models, protecting pretrained fidelity whenever the training set is tiny.","Not claimed by the paper: the dependence on five training images implies a concrete stress test, namely that corrupting or occluding one first-view image should worsen opposite-view consistency, and the paper's limitation section predicts exactly this failure mode.","Not claimed by the paper: the resample-and-merge loop could transfer to other autoregressive multi-view generation tasks, such as video inpainting where the \"known\" region is a previous frame rather than a painted mesh surface."],"forward_implications":["With $R$-fold resampling, each view needs roughly 30 denoising steps instead of the 250 steps of the underlying inpainting sampler, so the consistency gain does not come from brute-force compute per view.","Fine-tuning on five self-generated resampled images removes the need for an external dataset of 3D meshes, textures, or text descriptions; the object's own first views supply the training data.","Because the preservation loss keeps the fine-tuned model close to the pretrained predictor, the method avoids the catastrophic forgetting that naive few-image fine-tuning produces, as shown in the paper's ablation study.","Adaptive camera insertion means texture synthesis no longer requires manually tuned viewpoint sets; the method places an intermediate view whenever the visible painted fraction is too small.","On the reported Objaverse subset, the full method lowers LPIPS from 38.89 to 37.89 and FID from 30.924 to 29.998 compared to the no-training baseline, and beats the listed baselines on both metrics."],"supporting_citations":[{"why":"Supplies the resampling scheme that TexTailor adapts from 2D inpainting to the non-Markovian DDIM process.","marker":"(Lugmayr et al., 2022)"},{"why":"Defines the four-region texture synthesis procedure and the Objaverse evaluation subset that TexTailor builds on and compares against.","marker":"(Chen et al., 2023a)"},{"why":"Provides the TEXTure baseline that TexTailor compares against for view-consistency and quality.","marker":"(Richardson et al., 2023)"},{"why":"Provides the Latent-Paint baseline that TexTailor compares against in the quantitative evaluation.","marker":"(Metzer et al., 2023)"},{"why":"Provides the Paint-it baseline that TexTailor compares against in the quantitative evaluation.","marker":"(Youwang et al., 2024)"},{"why":"Provides the depth-conditioned diffusion architecture that TexTailor fine-tunes with the preservation loss.","marker":"(Zhang et al., 2023)"},{"why":"Gives the non-Markovian DDIM sampler that lets TexTailor run resampling in 30 steps.","marker":"(Song et al., 2020)"},{"why":"Supplies the Objaverse dataset used for the quantitative evaluation and comparisons.","marker":"(Deitke et al., 2022)"},{"why":"Supplies the LPIPS perceptual metric used to measure view-consistency across rendered pairs.","marker":"(Zhang et al., 2018)"},{"why":"Supplies the FID metric used to evaluate quality and diversity of the synthesized texture distribution.","marker":"(Heusel et al., 2017)"}],"fun_headline_variants":["Resampling stops texture drift across 3D views","Repeated resampling keeps 3D textures consistent","TexTailor: resampling for view-consistent textures","Suppress texture drift with resampling and fine-tuning","Consistent 3D texturing via resampling and loss preservation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that five images rendered near the first viewpoint teach the model the object's appearance at every other angle, even though the paper's limitation section concedes that texture quality \"heavily relies\" on those five images and that bad training images can degrade output.","fun_headline_variants_meta":{"raw":{"variants":["Resampling stops texture drift across 3D views","Repeated resampling keeps 3D textures consistent","TexTailor: resampling for view-consistent textures","Suppress texture drift with resampling and fine-tuning","Consistent 3D texturing via resampling and loss preservation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000679,"raw_usage":{"total_tokens":3103,"prompt_tokens":980,"completion_tokens":2123,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2043}},"tokens_in":596,"tokens_out":2123,"duration_ms":18297,"temperature":1.0,"reasoning_tokens":2043,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:21:39.910037+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a mesh with strong front/back asymmetry using a deliberately corrupted first-view training image, then compare the texture at the opposite viewpoint to the uncorrupted run: if the opposite view does not visibly drift or degrade, the fine-tuning step is more robust than the paper's stated five-image dependence.","supporting_citations":[{"cited_title":"Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering","cited_arxiv_id":null,"evidence_quote":"Provides the Paint-it baseline that TexTailor compares against in the quantitative evaluation."},{"cited_title":"Gans trained by a two time-scale update rule converge to a local nash equilibrium","cited_arxiv_id":null,"evidence_quote":"Supplies the FID metric used to evaluate quality and diversity of the synthesized texture distribution."}],"review_version":1}