{"id":"9b5ce39a-ef17-4aa1-b72d-b74b510b1810","arxiv_id":"2411.14740","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A feed-forward diffusion model generates high-resolution, globally coherent UV textures for arbitrary 3D meshes from text or image prompts in about 10 seconds.","lead":"TEXGen trains a 700-million-parameter diffusion model that paints UV texture maps onto 3D meshes from a text prompt or a single image, without per-object optimization. It brings feed-forward speed, about 10 seconds per mesh, to a task that previously required minutes of iterative optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Image conditioning requires pose- and shape-aligned views; arbitrary photo inputs break the pixel-projection path, narrowing the claimed single-view-image capability.","rationale":"The reader's weakest_assumption identifies the pose-alignment requirement as the most load-bearing concern, and I agree. The central claim in the abstract and conclusion presents the model as guided by single-view images without stating the alignment condition. Since the pixel-projection path in Section 4.2 is the only mechanism for injecting local pixel information from the image, and it requires known camera pose and shape correspondence, the model's image-conditioning capability is valid only for aligned views. This is a scope limitation that should be reflected in the claim, not a flaw in the model's internal consistency. The authors explicitly disclose it in Appendix A.3, which is a point in the paper's favor, but the abstract should still be qualified or the method should be shown to handle arbitrary photos. Other concerns, such as CFG-weight tuning on the test set and the comparison against methods not trained on Objaverse, are valid but secondary: they affect the strength of the quantitative 'state-of-the-art' subclaim, whereas the alignment issue affects the scope of the core announced capability. A CONDITIONAL verdict is appropriate: the paper's architectural contribution and large-scale training are credible, but a precise restatement of the image-conditioning scope (or evidence that the model works without alignment) is needed before the abstract's claim can be accepted as stated.","tokens_in":17388,"tokens_out":10537,"duration_ms":106532,"concrete_test":"Using the released code, take 100 Objaverse test meshes and, for each, generate a texture with an aligned condition image (a rendering of that mesh from a known viewpoint). Then generate textures with a misaligned condition: a photograph of a different object instance from the same category (e.g., a different chair photo for a chair mesh), with the text prompt set to empty. Compare the two sets of outputs by (a) measuring CLIP similarity between the input photo and multi-view renderings of the generated texture, and (b) running the same user-preference protocol used in the paper. If the misaligned condition significantly reduces identity preservation or user preference relative to the aligned condition, the limitation is confirmed and the abstract's image-guided claim must be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim advertises generation 'guided by text prompts and single-view images,' but the image-conditioning path is implemented by projecting image pixels onto the mesh surface (Section 4.2), which requires the input image to be pose-aligned and shape-aligned with the target mesh. This assumption is explicitly acknowledged in Appendix A.3: 'the condition images used during the training of our model are pose-aligned and shape-aligned.' For arbitrary user photos, the pixel projection is undefined or incorrect, so the only remaining image signal is a global CLIP embedding. That global embedding can provide semantic style but cannot transfer pixel-level appearance from the photo to corresponding surface regions. Consequently, the advertised single-view-image guidance works in the restricted setting of rendered, aligned views (or ControlNet-generated views from the mesh's own depth), but not for the common use case of texturing a mesh from a casual photograph of another object. The abstract and conclusion do not qualify this scope, so the central claim is broader than what the method delivers. This is a true limitation, not an internal inconsistency, and the authors deserve credit for stating it, but it directly affects the claimed capability and should be part of any assessment of the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TEXGen, a 700M-parameter diffusion model that generates 1024x1024 UV texture maps for arbitrary input meshes conditioned on text and a single view. The core technical novelty is a hybrid 2D-3D denoising block that interleaves UV-space convolutions (for local detail) with sparse point-cloud attention (for global 3D consistency). The model is trained on 120,000 Objaverse meshes with baked UV maps and evaluated on 400 held-out objects. Experiments report FID/KID improvements over TEXTure, Text2Tex, and Paint3D (e.g., FID 34.53 vs 43.55), a user-study preference of 69.3%, and MLLM-based text-alignment scores, plus ablations of the hybrid block and CFG weight. The paper also demonstrates zero-shot applications in inpainting, sparse-view completion, and scene texturing.","tokens_in":17597,"tokens_out":9230,"duration_ms":83243,"significance":"If the results hold, TEXGen is a valuable step: it demonstrates that a large feed-forward diffusion model can produce globally consistent, high-resolution textures on general meshes without per-object test-time optimization, and its hybrid 2D-3D architecture is a plausible scalable pattern for UV-space generation. The empirical support is substantial: held-out evaluation, multiple metrics, ablations, and a public code link. The central claims are nevertheless stated more broadly than the method delivers because the image-conditioning path requires pose- and shape-aligned input views (Appendix A.3), and quantitative comparisons are given as point estimates without uncertainty, which is particularly relevant to the CFG-weight selection.","major_comments":[{"comment":"The paper advertises generation 'guided by text prompts and single-view images' (abstract) without noting that the image-conditioning path projects image pixels onto the mesh surface (Section 4.2) and is only valid for pose-aligned, shape-aligned inputs, as acknowledged in Appendix A.3: 'the condition images used during the training of our model are pose-aligned and shape-aligned.' For an arbitrary photograph of an object, the projection is undefined and the image signal reduces to a global CLIP embedding that cannot transfer pixel-level appearance. The abstract, introduction, and conclusion should therefore qualify the image-conditioned capability (e.g., 'pose-aligned single-view images' or 'rendered views of the target mesh') and state the scope limitation up front. As written, the central claim of single-view-image guidance is broader than what the method delivers.","section":"Abstract and Section 4.2 / Appendix A.3"},{"comment":"The choice of CFG weight omega=2.0 for all main results is not clearly supported by the reported metrics. In Table 4, FID is lowest at omega=2 (34.53) but KID is lower at omega=3 (11.71 vs 11.94); the text says the weight 'around 2-3 is optimal,' yet the main experiments fix omega=2.0. No confidence intervals, bootstrap errors, or significance tests are given, so it is unclear whether the differences among omega in {2,3} are meaningful. If the guidance weight is selected on the evaluation set, this also risks optimistic bias. Please report uncertainty on FID/KID (e.g., bootstrap over the 400 test objects) and, ideally, select hyperparameters on a separate validation split or justify the choice with both metrics.","section":"Section 5.3, Table 4"},{"comment":"All quantitative comparisons are reported as point estimates without error bars or statistical tests. While the FID gap between TEXGen and Paint3D (34.53 vs 43.55) is large, the user-study result (69.3% preference from 423 responses) and the MLLM scores would be more convincing with confidence intervals or a significance test. Without these, the claimed state-of-the-art status rests on single seeds and standard benchmark implementations, which is fragile for generative models. Please add variance estimates, at least for the main comparison, or discuss why they are not applicable.","section":"Section 5.1, Tables 1 and 2"}],"minor_comments":[{"comment":"In the bullet on serialized attention, 'sterilized codes' should read 'serialized codes.'","section":"Section 4.2"},{"comment":"There is a typo in the caption ('feat u ring' should be 'featuring'), and the figure references mix 'fig.' and 'Fig.'; please standardize.","section":"Figure 5 caption"},{"comment":"The text states 'we compare our method with four representative state-of-the-art methods' but lists and shows only three baselines (TEXTure, Text2Tex, Paint3D). Please correct the count.","section":"Section 5.1 and Figure 6"},{"comment":"The sentence 'We use five stages to construct our network (i.e., four downsampling and four upsampling stages)' is ambiguous; clarify the number of resolution levels and how the down/up-sampling stages map to the five stages.","section":"Appendix A.1"},{"comment":"The ablation models are called Model A, B, C in the text but 'Hybrid block (A)', 'w/o point block (B)', 'w/o UV block (C)' in the table; please use consistent labels.","section":"Section 5.3 and Table 3"}],"recommendation":"major_revision","confidential_remarks":"This is a solid systems contribution and the limitation in Appendix A.3 is honestly disclosed, but the abstract overstates the image-conditioning capability. If the authors qualify the claims and add uncertainty quantification, I would support acceptance. The 'first feed-forward model' claim should also be checked against recent concurrent work (e.g., GenesisTex) to ensure the novelty statement is accurate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TEXGen is a legitimate step forward for mesh texturing. It trains a 700M-parameter diffusion model directly in UV space and generates 1024x1024 texture maps for general objects in a single feed-forward pass. As far as I can tell, that combination—direct UV-space learning, large scale, and no per-object optimization—is new. The hybrid 2D-3D block is the right kind of architectural idea: 2D convolutions keep high-frequency detail, sparse point-cloud attention restores the 3D neighborhood relationships that UV islands break. The ablations support it: dropping the point block hurts global consistency, dropping the UV block decimates high-frequency detail (FID 94 vs 70). The quantitative comparison against TEXTure, Text2Tex, and Paint3D is the best evidence, with a 9-point FID gap and roughly 2x faster runtime.\\n\\nThe largest caveat is the image-conditioning path. The pixel-projection mechanism in Section 4.2 requires the input image to be pose- and shape-aligned with the target mesh. For an arbitrary user photo, projection is undefined or wrong; the only remaining signal is a global CLIP embedding, which cannot transfer pixel-level appearance. The authors acknowledge this in Appendix A.3, which is honest, but the abstract's wording oversells it. Text-to-texture works and is useful because ControlNet synthesizes an aligned view from depth. For casual-photo texture transfer, this paper does not deliver.\\n\\nOther issues are minor and fixable. There are no error bars or significance tests on the FID/KID numbers, the CFG weight is selected on the evaluation metric, and the ablations use a reduced model on the house category. None of these sink the central claim. Also, the 'first feed-forward model' claim is slightly broad, but it is defensible if scoped to direct UV texture generation for arbitrary meshes.\\n\\nThe paper releases code, which is a plus. I'd send it to peer review and, if I were refereeing, ask for a scope qualifier on image conditioning and error bars if feasible. It deserves a serious referee.","headline":"A real and useful feed-forward texture generator whose single-view-image claim needs a scope qualifier, but the core contribution holds up.","tokens_in":18146,"tokens_out":3028,"would_cite":true,"duration_ms":29097,"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":"TEXGen trains a 700-million-parameter diffusion model that generates high-resolution UV texture maps for 3D meshes directly in one feed-forward pass, guided by text prompts or a single aligned image and without per-object test-time…","keywords":["texture generation","diffusion model","UV texture map","3D mesh","feed-forward generation","hybrid 2D-3D network","text-to-texture","image-guided generation"],"falsifier":"Render the generated texture from two viewpoints that look at neighboring surface patches belonging to different UV islands; if colors at the shared 3D seam differ strongly, the claimed global 3D coherence from the hybrid attention fails.","tokens_in":17169,"feed_emoji":"🎨","tokens_out":10156,"duration_ms":81938,"temperature":0.7,"pith_summary":"This paper claims that high-resolution texture maps for arbitrary 3D meshes can be generated directly by a single large diffusion model in one feed-forward pass, guided by a text prompt or a single viewpoint image, without the per-object test-time optimization used by prior texturing pipelines. The authors train a 700-million-parameter model on 120,400 cleaned mesh-texture pairs whose UVs were re-unfolded into a single atlas, and they report lower FID/KID and roughly an order-of-magnitude faster inference than TEXTure, Text2Tex, and Paint3D. If true, this turns mesh texturing from an optimization problem into a foundation-model problem, enabling training-free inpainting, sparse-view completion, and text-to-texture synthesis on general objects.","feed_headline":"First feed-forward diffusion model textures 3D meshes directly","feed_subtitle":"A 700M-parameter model turns text or one aligned photo into coherent 1024×1024 UV maps, no test-time optimization.","key_machinery":"The load-bearing component is the hybrid 2D-3D denoising block, which interleaves 2D convolutions on the UV map with 3D point-cloud attention. The UV head extracts local high-frequency detail efficiently on the grid; rasterization then lifts the features onto the mesh surface, where serialized attention (using space-filling-curve ordering such as Hilbert curves) and a modified conditional positional encoding (sCPE) propagate information across disconnected UV islands, restoring global 3D coherence. Condition modulation via learned scale-and-shift and gated fusion, following the DiT design, injects the text/image embedding; this block is what makes direct high-resolution UV diffusion scalable.","core_discovery":"The paper's central claim is that the UV texture map itself can be the direct learning target for a large-scale generative model, and that a hybrid 2D–3D diffusion network can learn this distribution at 1024×1024 resolution. The authors state that this is the first time a large diffusion model has been trained to generate texture maps end-to-end in a feed-forward manner, without additional stages or test-time optimization. The model is conditioned on both CLIP image and text embeddings, plus a partial texture map obtained by projecting the image pixels onto the surface, and it is trained with a v-prediction diffusion loss augmented by an LPIPS loss on multi-view renderings. At inference, classifier-free guidance with a weight near 2.0 produces the best FID, and the same network supports text-only synthesis, texture inpainting, and completion from sparse views without fine-tuning.","pith_inferences":["A natural next step the paper leaves implicit is replacing the pixel-projection alignment requirement with dense cross-attention, which would allow texture transfer from arbitrary photographs, not just pose- and shape-aligned ones.","The hybrid 2D-3D block is not tied to color textures; the same interleaving of UV convolutions and surface attention could be applied to other per-surface signals, such as normal maps, roughness maps, or PBR material stacks.","The reported speed advantage, if it holds at scale, suggests feed-forward texture models may become the default inside game and animation pipelines, where the cost of per-asset optimization currently dominates.","One testable extension: measure whether the model's cross-island consistency degrades gracefully as UV island count increases; if it does not, the serialized-attention component would need stronger geometric grouping."],"forward_implications":["Texture generation for general 3D objects becomes a single forward pass of about 10 seconds on an A100, replacing 80–344 seconds of per-object optimization in prior methods.","The same pretrained model can be used without fine-tuning for text-guided inpainting of partial texture maps, for completing textures from two or more sparse views, and for pure text-to-texture synthesis when combined with a depth-conditioned image generator.","Because the model is trained directly on 3D data and full UV maps, it avoids view-dependent artifacts such as the Janus problem.","The architecture's scalability suggests further gains from larger datasets and model sizes, and the feed-forward nature makes it compatible with diffusion acceleration techniques such as consistency distillation."],"supporting_citations":[{"why":"Supplies the raw data source: over 800,000 Objaverse meshes, from which the 120,400 cleaned mesh-texture pairs are derived.","marker":"[Deitke et al. 2023]"},{"why":"Defines the prior UV-space diffusion approach that lacked end-to-end training and feed-forward inference, which the paper's novelty is contrasted against.","marker":"[Yu et al. 2023a]"},{"why":"Provides the base noise scheduler from Stable Diffusion, adapted with zero-terminal SNR for the diffusion process.","marker":"[Rombach et al. 2022]"},{"why":"Supplies the v-prediction parameterization used in the diffusion training objective.","marker":"[Salimans and Ho 2022]"},{"why":"Introduces classifier-free guidance, used at inference after random dropping of condition embeddings during training.","marker":"[Ho and Salimans 2022]"},{"why":"Provides the DiT-style condition modulation with scale-and-shift and gated fusion used in the hybrid blocks.","marker":"[Peebles and Xie 2023]"},{"why":"Supplies serialized attention and the conditional positional encoding (xCPE/sCPE) for efficient point-cloud attention in 3D space.","marker":"[Wu et al. 2023]"},{"why":"Defines the LPIPS perceptual loss applied to multi-view renderings as auxiliary supervision.","marker":"[Zhang et al. 2018]"}],"fun_headline_variants":["First diffusion model to generate mesh textures directly in UV space","Feed-forward diffusion: high-res 3D textures from text or a single image","TEXGen: one-pass UV texture generation with 700M-param diffusion","No test-time tricks: diffusion model paints 3D meshes from text or photo","Direct UV-space diffusion for instant texturing of 3D assets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The image-conditioning path assumes, as the paper's Appendix A.3 acknowledges, that input photos are pose-aligned and shape-aligned with the target mesh; with arbitrary user photos the pixel-projection step cannot be applied, so image-guided texture generation breaks down.","fun_headline_variants_meta":{"raw":{"variants":["First diffusion model to generate mesh textures directly in UV space","Feed-forward diffusion: high-res 3D textures from text or a single image","TEXGen: one-pass UV texture generation with 700M-param diffusion","No test-time tricks: diffusion model paints 3D meshes from text or photo","Direct UV-space diffusion for instant texturing of 3D assets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000839,"raw_usage":{"total_tokens":3641,"prompt_tokens":910,"completion_tokens":2731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":2632}},"tokens_in":526,"tokens_out":2731,"duration_ms":20519,"temperature":1.0,"reasoning_tokens":2632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:56:59.355948+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the generated texture from two viewpoints that look at neighboring surface patches belonging to different UV islands; if colors at the shared 3D seam differ strongly, the claimed global 3D coherence from the hybrid attention fails.","supporting_citations":[],"review_version":1}