{"id":"87d86fa2-4e72-4cae-87ff-8dadc91952ff","arxiv_id":"2506.02846","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PBR-SR super-resolves PBR texture maps (albedo, roughness, metallic, normal) in a zero-shot way by optimizing textures so differentiable renderings match super-resolved multi-view renderings from a pretrained image SR model.","lead":"PBR-SR is a method that increases the resolution of physically based rendering (PBR) textures on 3D meshes using a pretrained image super-resolution model, without retraining. A generalist reader might care because it could upgrade old game and 3D assets to high detail while preserving material properties for relighting.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Same-light quantitative evaluation leaves the relighting benefit unverified; a held-out environment-map rendering test is the decisive check that should be run before the relighting claim is accepted.","rationale":"The reader's verdict is CONDITIONAL with medium correctness risk, and I agree. I considered whether the more fundamental concern is DiffBIR hallucinating high-frequency detail. That would be a problem, but the robust pixel-wise loss (Eqs. 3-5) is explicitly designed to downweight multiview-inconsistent hallucinations, and the identity loss (Eq. 6) anchors the result to the LR input; without a concrete failure case this is a risk rather than a demonstrated weakness. The relighting issue is stricter: the paper's own protocol uses the same environment for optimization and evaluation, and the optimization objective has no term that prevents baking the fixed environment into the material maps. The qualitative relighting figures do not include GT or baseline comparisons, so they cannot establish transfer. A single quantitative novel-lighting experiment would settle the matter. Because the concern is about an unsupported extension of the claim rather than an observed failure, the verdict should remain CONDITIONAL: accept only with the novel-lighting evaluation added, or with the relighting claim explicitly scoped to the optimization environment. The reader identified the same same-light issue, so I leave the conditional verdict unchanged.","tokens_in":13684,"tokens_out":5949,"duration_ms":73570,"concrete_test":"Run a held-out-lighting evaluation. Take the 16-mesh benchmark, optimize PBR-SR with the default environment used in the paper, then render the optimized maps under three unseen environment maps (e.g., Billiard Hall, De Balie, Beach Parking from Fig. 8) and compare to GT renderings under those same lights using PSNR, SSIM, and LPIPS over the same 240 novel views. Include the strongest baselines (DiffBIR, CAMixerSR-FT, Paint-it SR) under the same novel lights. Report per-mesh mean and standard deviation. If PBR-SR's margin over the best baseline shrinks or reverses under novel lighting, the same-light evaluation protocol is the source of the claimed gain and the relighting claim must be weakened. If the margin survives, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes 'supporting advanced applications such as relighting,' but the quantitative support for this is missing. Section 4.1 states that 'Unless explicitly specified, the same environment lighting setup is used for both optimization and evaluation,' and Table 1's rendering PSNR is computed under that same environment. During optimization (Eq. 3), the pseudo-GT images are DiffBIR super-resolutions of LR renderings under that fixed environment map. Moving the camera around the object does not vary the incident illumination at each surface point: for a fixed environment, diffuse irradiance at a surface point is view-independent, and only the specular lobe changes with view. Thus the multi-view optimization does not by itself separate material from lighting. The PBR consistency loss (Eq. 6) only forces the pooled HR texture to match the LR texture; it does not constrain the SR maps to be light-invariant. Consequently, high rendering PSNR in Table 1 can be achieved by absorbing shading into albedo or by encoding the fixed light direction into normal/roughness, while still passing the Eq. 6 identity check. The only relighting evidence offered is qualitative (Figs. 5, 6, 8) with no ground-truth comparison or baseline comparison under novel lights. This is not an internal inconsistency, but it is a load-bearing unsupported assertion: if the optimized maps overfit to the optimization environment, the central relighting claim fails even though the same-light SR numbers remain high.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PBR-SR, a zero-shot method for super-resolving PBR texture maps (albedo, roughness, metallic, normal) on meshes. It initializes high-resolution textures by combining a pretrained image SR model (DiffBIR) on the albedo map with bicubic upsampling of the ARM and normal maps. It then renders the mesh from multiple viewpoints, uses DiffBIR to super-resolve those renderings into pseudo-ground-truth images, and iteratively optimizes the PBR textures so that differentiable renderings match those pseudo-GTs. A robust per-pixel weighting map is introduced to downweight unreliable regions, and PBR consistency and TV losses keep the SR textures faithful to the LR input. Experiments on a self-collected set of 16 PBR meshes report PSNR for albedo, roughness, metallic, normal, and renderings, and the authors claim consistent state-of-the-art performance and support for relighting.","tokens_in":14022,"tokens_out":3833,"duration_ms":43950,"significance":"If its claims hold, PBR-SR would be a practical contribution: it requires no training data, reuses off-the-shelf image SR priors, and directly outputs material maps that can be relit. The formulation is clean and the ablations show that each loss component contributes positively. However, the evidence in the paper is substantially thinner than the claims. The quantitative evaluation is PSNR-only on 16 meshes with no error bars or significance tests, the relighting claim is not tested quantitatively under held-out lighting, and the pseudo-GT supervision comes from the same DiffBIR model used for initialization. These are load-bearing gaps for the core claims of state-of-the-art performance and relighting support, not mere presentation issues.","major_comments":[{"comment":"The relighting claim is not supported by the evaluation. Sec. 4.1 states that 'Unless explicitly specified, the same environment lighting setup is used for both optimization and evaluation,' and Table 1's rendering PSNR is computed under that same environment. Under a fixed environment map, diffuse irradiance at a surface point is view-independent; only the specular term varies with view. Therefore the multi-view optimization in Eqs. (2)-(3) does not by itself separate material properties from lighting, and the PBR consistency loss in Eq. (6) only enforces pooled LR fidelity, not light invariance. High same-light rendering PSNR can be achieved by absorbing shading into albedo or encoding the fixed light direction into normal/roughness. The only relighting evidence is qualitative (Figs. 5, 6, 8) with no ground-truth or baseline comparison under novel lights. I recommend a held-out environment-map evaluation with quantitative metrics (e.g., PSNR/LPIPS against GT renderings under several unseen environment maps) and comparison to baselines before the 'supporting advanced applications such as relighting' claim is accepted.","section":"Sec. 4.1 and Table 1"},{"comment":"The state-of-the-art claim rests on a very thin evaluation. The dataset is a self-collected set of only 16 meshes, and Table 1 reports mean PSNR without standard deviations, per-mesh distributions, or significance tests. Several margins are small (e.g., Metallic 31.889 vs. 30.536 for HAT; Normal 29.088 vs. 28.237 for Paint-it SR), so the claim that PBR-SR 'consistently outperforms all baselines' is not statistically supported. In addition, hyperparameters (loss weights, DiffBIR adaptation, rendering resolution) appear to have been selected on the same collection, with Table 5 explicitly tuning resolution on the Table Clock mesh, and no separate validation split is described. The paper should report per-mesh results with uncertainty, run significance tests, and ideally evaluate on a larger or independently collected set.","section":"Sec. 4.1 and Table 1"},{"comment":"The supervision loop may limit the method's ability to recover genuine high-frequency material detail. The pseudo-GT images are produced by DiffBIR super-resolving renderings of the initial SR textures, and the albedo initialization is also DiffBIR-based. This creates a self-distillation loop: the optimization can only inject details that DiffBIR hallucinates or extrapolates, and those hallucinations can be baked into the PBR maps while Eq. (6) still passes because it only checks averaged LR fidelity. The paper does not analyze what happens when DiffBIR produces incorrect or inconsistent high-frequency content, nor does it compare against using a different SR prior for initialization versus pseudo-GT generation. I would like to see an explicit discussion or ablation that separates the contribution of the LR fidelity constraint from the prior's hallucinated details, for example by measuring how much of the PSNR gain comes from channels where DiffBIR is known to be unreliable.","section":"Sec. 3.4 and Sec. 3.5"}],"minor_comments":[{"comment":"The notation in the text is reversed: the predicted rendering is called I^SR_i and the target pseudo-GT is called I^HR_i, but Eq. (3) uses I^SR_i as the target and I^HR_i as the prediction. Please align the notation.","section":"Sec. 3.5.1"},{"comment":"The evaluation metrics paragraph contains a typo: 'A higher PNSR value is better' should read 'PSNR'.","section":"Sec. 4.2"},{"comment":"The ablation on rendering resolution is reported only on a single mesh (Table Clock). Please clarify that this is a single-mesh study and avoid drawing general conclusions from it without additional meshes.","section":"Supplementary Table 5"},{"comment":"The supplementary material contains a placeholder reference 'Fig. X' in the discussion of the robust pixel-wise loss; this should be fixed to the actual figure number.","section":"Supplementary Fig. 5"},{"comment":"The handling of the ARM map is unclear: if AO is unavailable, an empty map is allocated in the red channel. Please specify how this placeholder interacts with the PBR consistency loss and the renderer, since an empty AO channel may introduce unintended shading.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a useful problem and the core optimization idea is plausible, but the current evidence is not yet sufficient for the claims made. The same-light evaluation and the absence of quantitative relighting tests are the main blockers. I would not reject the paper; the missing experiments are well-defined and within the scope of a revision. The authors should also be encouraged to share code and per-mesh results to strengthen reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea is real and worth building on—zero-shot PBR texture SR by optimizing textures against a frozen 2D SR prior through differentiable PBR rendering—but the quantitative case is much thinner than the abstract lets on, and the relighting claim is currently unsupported. I'd send it to reviewers, but they'll need to push for proper evaluation.\n\nWhat's actually new: their literature review finds only [19] for mesh texture SR, and that method handles baked appearance with finetuning. This is the first zero-shot PBR-map SR, and the pipeline—DiffBIR init in UV space, multi-view render-based refinement, robust per-pixel weighting, and a texture-space consistency loss—is a sensible combination. The ablations show each piece contributes, and Table 1's margins over the supervised CAMixerSR-FT baseline are a good sign. Qualitatively, the renderings are visibly sharper and cleaner.\n\nNow the soft spots, in order of importance. First and most load-bearing: relighting is claimed but never measured under novel lighting. Section 4.1 states that unless explicitly specified, the same environment lighting is used for optimization and evaluation. That is exactly the test that cannot separate material from lighting. For a fixed environment, diffuse irradiance at a surface point is view-independent, so multi-view supervision does not by itself constrain light-invariance. The PBR consistency loss only forces the pooled SR texture to match the LR texture; it does not stop the optimizer from baking the environment's key light direction into normal or roughness. The high rendering PSNR can therefore be achieved while the SR maps overfit to the optimization environment. The only relighting evidence is qualitative—Figs. 5, 6, 8—with no ground-truth comparison or baselines under novel lights. The decisive check is a held-out environment-map rendering test with PSNR versus GT, run for both the method and baselines under the same novel lights. Without that, the relighting claim should not stand.\n\nSecond, the evaluation is thin in other ways: 16 meshes, PSNR only, no error bars or significance tests. The margins over CAMixerSR-FT could be within noise. No code or data release, and the many loss weights and DiffBIR settings appear tuned on the same collection. Third, the pseudo-GT loop: DiffBIR is used to initialize the textures and also to generate pseudo-GT from renderings of the same LR inputs. That is a self-distillation setup, not necessarily a fatal flaw—the consistency loss keeps outputs anchored—but the supervision is not independent of the prior. Reviewers should ask how much of the reported PSNR is reproducing DiffBIR's own hallucinations.\n\nBottom line: the method is clever, clearly written, and honestly positioned in the literature. The stress-test note about same-light evaluation is correct and should be fixed before acceptance. I'd bring this to reading group and cite it if I worked on mesh texture SR or zero-shot PBR optimization. Verdict: deserves a serious referee, conditional on a held-out-lighting experiment and released code/data.","headline":"Clever zero-shot PBR texture SR that is genuinely new but needs a held-out-lighting evaluation before the relighting claim can be trusted.","tokens_in":14545,"tokens_out":2155,"would_cite":true,"duration_ms":25542,"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":"PBR-SR claims that zero-shot super-resolution of PBR material maps is achievable by distilling 2D image super-resolution priors through differentiable rendering, with top PSNR on all five evaluated metrics.","keywords":["physically based rendering","texture super-resolution","zero-shot","differentiable rendering","diffusion image priors","multi-view consistency","material maps","relighting"],"falsifier":"Render the optimized material maps under an environment map that was not used during optimization and compare PSNR against ground truth; if the margin over a direct texture-super-resolution baseline disappears, the relighting claim is falsified. A second probe is to feed synthetic low-resolution textures generated from ground-truth maps and check whether optimized normal and roughness maps stay close to ground truth under changing lights instead of drifting to compensate for baked shadows.","tokens_in":13513,"feed_emoji":"🎨","tokens_out":9128,"duration_ms":81612,"temperature":0.7,"pith_summary":"PBR-SR claims that low-resolution physically based rendering textures—albedo, roughness, metallic, and normal maps—can be super-resolved to high resolution without paired training data by distilling a pretrained 2D natural-image super-resolution model through differentiable rendering. The method initializes high-resolution maps from the low-resolution input, renders the mesh from many viewpoints, treats the 2D super-resolved renderings as pseudo-ground truths, and iteratively optimizes the shared PBR maps so that differentiable renderings match those pseudo-ground truths while staying faithful to the original low-resolution textures. If correct, this gives a practical route to upgrading legacy and AI-generated 3D assets, with output material maps remaining renderable under new lighting rather than being baked images. The paper reports the highest PSNR across albedo, roughness, metallic, normal, and rendering metrics, beating direct super-resolution baselines, an optimization-based baseline, and a supervised baseline fine-tuned on PBR data.","feed_headline":"PBR textures get 4x super-resolution without training data","feed_subtitle":"Iterative differentiable rendering distills 2D image super-resolution priors into albedo, roughness, metallic, and normal maps.","key_machinery":"The carrying mechanism is the iterative optimization loop around a differentiable PBR renderer. A per-image, per-pixel weighting map is optimized alongside the textures to downweight unreliable pseudo-ground-truth regions such as shadows, specular highlights, and view-inconsistent details; the robust loss uses the squared weights to modulate a normalized mean-squared error. The PBR consistency loss is the identity anchor: it average-pools each refined high-resolution map back to the low-resolution grid and compares it to the input with L1 plus SSIM, so the optimized textures cannot drift away from the low-resolution input while they absorb new detail. A total-variation term suppresses artifacts. The pretrained blind image restoration diffusion model supplies both the albedo initialization and the pseudo-ground-truth renderings.","core_discovery":"The central claim is that 2D image super-resolution priors can be transferred into 3D PBR texture space by a render-and-distill loop. From multiple viewpoints, the current high-resolution textures are rendered with a differentiable PBR rasterizer, and the same viewpoints are rendered with a conventional rasterizer and upscaled by a pretrained blind image restoration diffusion model to create pseudo-ground-truth images. Optimizing a robust pixel-wise loss between the two forces the texture maps to absorb the added high-frequency detail, while a PBR consistency loss that downsamples the refined maps back to low resolution preserves the input's structure and material identity. The result is a zero-shot super-resolution method that works on artist-designed and AI-generated meshes alike and supports relighting, which direct image super-resolution on texture maps cannot do.","pith_inferences":["Inference: the quantitative evaluation of rendering quality uses the same environment lighting for optimization and evaluation, so the relighting claim should be tested with held-out lighting before relying on it.","Inference: the jointly optimized per-pixel weighting maps encode where the 2D prior is uncertain, and could be reused as confidence maps for material editing or for choosing better viewpoints in a second optimization pass.","Inference: the same render-and-distill loop should extend to other material properties such as clearcoat, anisotropy, or displacement whenever a differentiable renderer and a pseudo-ground-truth generator exist for those channels."],"forward_implications":["Any pretrained 2D image super-resolution model can be plugged into the same loop, so future super-resolution models can improve PBR-SR without retraining or architectural changes.","PBR-SR upgrades both artist-designed legacy assets and AI-generated PBR textures, though generated textures with weak structural cues limit how much detail can be recovered.","Because the output is material maps rather than baked appearance, the upsampled textures can be relit under new environments, which direct texture super-resolution and baked-texture super-resolution cannot do.","The robust pixel weighting allows the optimization to tolerate multi-view inconsistencies and lighting artifacts in the pseudo-ground truths, which is the main reason view-based super-resolution approaches fail."],"supporting_citations":[{"why":"Supplies the pretrained blind image restoration diffusion model used to produce pseudo-ground-truth renderings and initialize the albedo map.","marker":"[22]"},{"why":"Supplies the differentiable mesh rasterizer used to render current textures and back-propagate gradients into the PBR maps.","marker":"[17]"},{"why":"Defines the Cook-Torrance BRDF and rendering equation that the differentiable renderer evaluates, connecting PBR maps to images.","marker":"[11]"},{"why":"Provides part of the high-quality PBR mesh and texture collection used to build the evaluation set.","marker":"[1]"},{"why":"Provides part of the high-quality PBR mesh and texture collection used to build the evaluation set.","marker":"[2]"},{"why":"Provides the PBR mesh collection used for evaluation, the environment maps for relighting tests, and the data used to fine-tune the supervised baseline.","marker":"[3]"},{"why":"Provides the CAMixerSR image super-resolution baseline and the architecture fine-tuned on PBR data as the supervised CAMixerSR-FT comparison.","marker":"[32]"},{"why":"Provides Paint-it, whose optimization framework is adapted into the optimization-based baseline Paint-it SR.","marker":"[34]"},{"why":"Supplies the SwinIR transformer image super-resolution baseline compared in the main results table.","marker":"[20]"}],"fun_headline_variants":["Zero-shot PBR texture upscaling via 2D image priors","Render-and-distill boosts PBR maps 4x without training","Multi-view render-distill upsamples PBR textures without retraining","PBR maps get 4x resolution using only pretrained 2D priors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pretrained 2D image enhancer produces upscaled example images whose fine detail is true material detail, not hallucination or baked lighting, and that this detail stays correct under lighting the optimizer never saw.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot PBR texture upscaling via 2D image priors","Render-and-distill boosts PBR maps 4x without training","Multi-view render-distill upsamples PBR textures without retraining","PBR maps get 4x resolution using only pretrained 2D priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00108,"raw_usage":{"total_tokens":4518,"prompt_tokens":943,"completion_tokens":3575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":3494}},"tokens_in":559,"tokens_out":3575,"duration_ms":22945,"temperature":1.0,"reasoning_tokens":3494,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:14:37.315639+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the optimized material maps under an environment map that was not used during optimization and compare PSNR against ground truth; if the margin over a direct texture-super-resolution baseline disappears, the relighting claim is falsified. A second probe is to feed synthetic low-resolution textures generated from ground-truth maps and check whether optimized normal and roughness maps stay close to ground truth under changing lights instead of drifting to compensate for baked shadows.","supporting_citations":[{"cited_title":"Modular primitives for high-performance differentiable rendering.ACM Transactions on Graphics, 39(6), 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable mesh rasterizer used to render current textures and back-propagate gradients into the PBR maps."},{"cited_title":"Torrance","cited_arxiv_id":null,"evidence_quote":"Defines the Cook-Torrance BRDF and rendering equation that the differentiable renderer evaluates, connecting PBR maps to images."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides part of the high-quality PBR mesh and texture collection used to build the evaluation set."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides part of the high-quality PBR mesh and texture collection used to build the evaluation set."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PBR mesh collection used for evaluation, the environment maps for relighting tests, and the data used to fine-tune the supervised baseline."},{"cited_title":"Camixersr: Only details need more ”attention”","cited_arxiv_id":null,"evidence_quote":"Provides the CAMixerSR image super-resolution baseline and the architecture fine-tuned on PBR data as the supervised CAMixerSR-FT comparison."},{"cited_title":"Paint-it: Text-to-texture synthesis via deep con- volutional texture map optimization and physically-based rendering","cited_arxiv_id":null,"evidence_quote":"Provides Paint-it, whose optimization framework is adapted into the optimization-based baseline Paint-it SR."}],"review_version":1}