{"id":"d6c54682-8334-4e64-9eeb-943bd8b35c2b","arxiv_id":"2508.19789","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"StableIntrinsic estimates albedo, roughness, and metallic maps from multi-view RGB images in a single diffusion step, achieving higher PSNR and lower MSE than prior multi-step diffusion methods.","lead":"This paper presents StableIntrinsic, a diffusion model that estimates surface materials (color, roughness, metalness) from photographs in one fast step instead of the usual dozens. The method runs roughly 50 times faster than the previous standard and produces more consistent and detailed material maps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline gains hinge on a self-selected test split and single-seed baseline numbers; without error bars and the released split, the claimed SOTA margins are not established.","rationale":"The reader's weakest_assumption focused on the self-constructed test split and the possibility of incomplete overlap exclusion. That is a real concern, but I would weight the missing error bars and selective abstract reporting as equally load-bearing, because IDArb is stochastic and the paper's own selling point is low variance relative to multi-step diffusion. A single-seed comparison can easily misstate the true margin. The paper is otherwise coherent: the method is clearly described, the ablation study supports the individual contributions (pixel-space loss, gradient-matching loss, DIN), and the one-step inference claim is technically accurate. The issue is not internal inconsistency but insufficient evaluation evidence for the headline quantitative claim. Thus the verdict should remain CONDITIONAL: the central claim is plausible but not yet established without the released split and multi-seed statistics.","tokens_in":12608,"tokens_out":3268,"duration_ms":41243,"concrete_test":"Release the exact 200-object Objaverse-Testing split (with object identifiers and exclusion logic) and re-run the comparison of StableIntrinsic vs IDArb using at least 10 inference seeds per object, reporting mean ± std for PSNR/SSIM/MSE. Also compute an ensembled IDArb baseline (average of the 10 predictions per object) and report its metrics. If ensembled IDArb meets or beats StableIntrinsic on the released split, the headline 'surpasses SOTA' would not hold under a practically meaningful comparison; if StableIntrinsic still wins, the concern is resolved. In addition, apply the same multi-seed protocol to Stanford-ORB and report the numbers that the abstract omits.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claim is that StableIntrinsic surpasses SOTA by 9.9% PSNR on albedo and 44.4%/60.0% MSE reductions on metallic/roughness. These numbers come from Objaverse-Testing, a 200-object split constructed by the authors (Sec. 5.1) because the official IDArb test split is unpublished. The split is not released, and the overlap-exclusion procedure ('exclude object indices present in the training set') is not independently verifiable. More importantly, IDArb is a stochastic multi-step diffusion method, so a single run may not represent its expected performance; the paper's own motivation is that such methods have high variance (Fig. 2), yet no variance or standard error is reported for any baseline or for StableIntrinsic in Tables 1-4. The abstract also selectively reports the split where the method wins, while Table 2 shows StableIntrinsic does not beat IDArb on Stanford-ORB PSNR (32.33 vs 32.57). Since the claimed superiority is quantitative and the magnitudes are moderate, the lack of error bars and a verified, shared evaluation split leaves the central 'surpasses current state-of-the-art' claim under-supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes StableIntrinsic, a one-step latent diffusion model for multi-view material estimation (albedo, roughness, metallic). The method fine-tunes a pre-trained multi-step diffusion model at the final time step T, applies pixel-space MSE and gradient-matching losses, and introduces a Detail Injection Network (DIN) that concatenates encoder features into the decoder to recover high-frequency detail. The authors report speedups of roughly 50x over IDArb and claim state-of-the-art quantitative results on synthetic data, with lower variance than multi-step baselines. Ablations support the value of optimization at t=T, the gradient-matching loss, and DIN.","tokens_in":12847,"tokens_out":4314,"duration_ms":46493,"significance":"If the claims hold, the paper makes a useful contribution: it is among the first to adapt one-step diffusion to material estimation, and the speed improvement (Table 4) is substantial. The pixel-space loss and DIN are plausible and the ablations in Table 3 support their individual contributions. The empirical gains on the self-constructed Objaverse-Testing split are large, and the qualitative results show clearer high-frequency detail than the baselines. However, the central 'surpasses SOTA' claim is currently under-supported by the evaluation protocol: the main benchmark is a non-released, self-selected split, and no variance/statistical reliability measures are reported for any method, despite the paper's focus on variance. The paper is technically plausible but needs stronger evaluation transparency before the headline claims can be accepted.","major_comments":[{"comment":"The central quantitative claim—a 9.9% PSNR improvement on albedo and large MSE reductions—is measured on Objaverse-Testing, a 200-object split constructed by the authors because IDArb's test split is unpublished. The split is not released, and the exclusion of training objects is not independently verifiable. More importantly, IDArb is stochastic, and the paper's own motivation is the high variance of multi-step diffusion, yet no error bars, standard deviations, or multiple-seed evaluations are reported for any baseline or for StableIntrinsic. A single run of a stochastic method is insufficient to support the claimed margins. Please release the exact split, repeat evaluations over several seeds, and report mean±std or confidence intervals for all methods on all datasets.","section":"Section 5.1, Tables 1-2"},{"comment":"The abstract states that StableIntrinsic 'surpasses the current state-of-the-art techniques' with the 9.9%/44.4%/60.0% improvements, but those numbers are only for Objaverse-Testing. Table 2 shows that on the real-world Stanford-ORB dataset, IDArb has higher albedo PSNR (32.57 vs 32.33). While the text acknowledges this, the abstract's blanket claim is misleading without qualification. Please either restrict the headline claim to the synthetic benchmark or include a caveat that on some real-world data the method does not lead on PSNR.","section":"Abstract and Table 2"},{"comment":"The main training loss is written as L_θ = E || -K - D(μθ(ε, z_c, T)) ||^2_2. With K the target material map, this objective drives D's output toward -K, which is inconsistent with Eq. (4) and with the reported behavior of the method. If this is a sign typo, please correct it to 'K - D(...)'; if not, the derivation is wrong and the loss cannot produce the presented results. This is a load-bearing equation and must be fixed.","section":"Eq. (5), Section 3.2"},{"comment":"The training objective for the Detail Injection Network f_φ is not specified. The text says 'we freeze all other network parameters and only optimize f_φ to enhance details,' but no loss function for f_φ is given. Since DIN is one of the three stated contributions, the missing loss is a reproducibility gap. Please provide the exact loss, data used, and whether it is the same pixel-space loss as for μθ.","section":"Section 4, 'Optimization'"},{"comment":"The claim that 'even if we remove the Gaussian noise input, it does not adversely affect the model's performance' is made without any experimental support. This claim is directly related to the paper's low-variance motivation and is easily testable. Either provide an ablation table showing the effect of removing the noise input or revise the text to state this as a hypothesis rather than a result.","section":"Section 3.4"}],"minor_comments":[{"comment":"Typo: 'PRB material' should be 'PBR material'.","section":"Section 3, first paragraph"},{"comment":"The notation 'εz_c' should be 'ε, z_c' (comma missing).","section":"Eq. (5)"},{"comment":"The 'scale-invariant PSNR' is never defined. Please clarify how it differs from standard PSNR and provide the formula.","section":"Section 5.1, Metrics"},{"comment":"For RGB↔X, metallic and roughness MSE are shown as '-', but no explanation is given. If the model does not output these maps, please state this in the caption or text.","section":"Table 1"},{"comment":"The variance maps would be more informative with a quantitative scale or normalization; as shown, the numbers (e.g., 42.52, 300.68) are hard to interpret without context.","section":"Figure 2"},{"comment":"Training uses 256×256 resolution but evaluation samples 4 viewpoints at 512×512. Please discuss potential resolution mismatch and whether the model was trained at higher resolution or upsampled.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper proposes a solid direction and the core method appears plausible, but the evaluation protocol is the key weakness. The self-constructed test split and the absence of any statistical reliability measures are not merely cosmetic; they are load-bearing for the 'surpasses SOTA' claim. I would be willing to recommend acceptance after the authors release the split, provide multi-seed variance numbers, fix the sign error in Eq. (5), and specify the DIN training objective. The novelty claim ('first one-step diffusion for material estimation') is defensible but should be checked against any concurrent work such as recent one-step dense prediction models adapted to material estimation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a solid engineering contribution that does what it says. One-step denoising for material estimation is a real practical win, especially the ~50x speedup and the visibly lower variance versus IDArb shown in Fig. 2. The pixel-space loss design, the gradient matching term for roughness/metallic, and the DIN detail-injection module are all sensible, and the ablations in Table 3 give each component clear credit. The method is coherent and the experiments, within their chosen scope, support the main qualitative claims.\n\nThe soft spots are exactly where the stress-test note lands. The headline numbers come from Objaverse-Testing, a 200-object split the authors constructed because the official IDArb test split is unpublished. That split is not released, and no error bars or variance statistics are reported for any method. Since the paper's own motivation is that stochastic multi-step methods have high variance, comparing against a single run of IDArb without variance is a real gap. The abstract also cherry-picks the split where the method wins; on Stanford-ORB, PSNR is lower than IDArb, and the explanation — pseudo-GT albedo is blurry — is plausible but doesn't convert a lower metric into a win. The missing code and data are practical barriers to verification.\n\nThat said, these issues don't kill the paper. The central idea — optimize a latent diffusion model at t=T in pixel space with task-specific losses and a detail-injection decoder — is sound, and the ablations back it. The MSE-loss/PSNR overlap is typical and not a real concern. The conclusion also states limitations honestly, including the highlight-injection failure and illumination sensitivity.\n\nWho gets value: anyone working on diffusion-based inverse rendering or fast dense prediction. The paper deserves a serious referee, but the review should demand the released test split, variance statistics for all baselines, and more careful real-world reporting. I'd accept it for peer review, not for publication as-is.","headline":"Useful engineering contribution with a real speed and variance win, but the headline SOTA numbers rest on a self-selected unreleased test split and single-seed baselines.","tokens_in":13382,"tokens_out":1974,"would_cite":true,"duration_ms":22436,"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":"StableIntrinsic claims that material estimation from multi-view images requires only one diffusion denoising step, with pixel-space losses and a detail-injection network preserving detail.","keywords":["material estimation","one-step diffusion","intrinsic decomposition","multi-view consistency","albedo","roughness","metallic","detail injection network"],"falsifier":"Take an independent set of objects with exact ground-truth albedo, roughness and metallic maps, or the official IDArb test split once released, and compare StableIntrinsic's metrics against IDArb's. If the albedo PSNR gain does not approach 9.9% or the metallic/roughness MSE reductions do not approach 44.4% and 60.0%, the central quantitative claim fails; the same protocol on glossy textured objects tests whether the Detail Injection Network's highlight leakage undermines the detail-preservation claim.","tokens_in":12480,"feed_emoji":"🎨","tokens_out":6565,"duration_ms":67439,"temperature":0.7,"pith_summary":"StableIntrinsic claims that diffusion-based material estimation from multi-view RGB images does not need dozens of denoising steps: a single forward pass at the pure-noise timestep, trained with pixel-space losses, can produce sharper albedo, roughness and metallic maps than the multi-step state of the art. The paper's central evidence is quantitative: a 9.9% higher albedo PSNR, a 44.4% lower metallic MSE and a 60.0% lower roughness MSE on a 200-object test set, with denoising roughly 50x faster than the IDArb baseline. It also argues that because only the timestep-T noise is involved, the model is nearly insensitive to the initial noise and therefore has much lower output variance than multi-step diffusion. If correct, this removes the main obstacles—slow inference and stochastic instability—that keep diffusion material estimators from being used in deterministic inverse-rendering pipelines.","feed_headline":"Diffusion model recovers materials in one step, 50x faster","feed_subtitle":"Pixel-space losses and a detail-injection network keep fine textures and low variance.","key_machinery":"The central object is the one-step v-parameterized latent diffusion model optimized only at t=T, where the input latent is pure Gaussian noise. Around it sit two additions: pixel-space losses chosen per material property (MSE for albedo, MSE plus gradient matching for roughness/metallic) and the Detail Injection Network, which feeds high-resolution encoder features into the decoder via Residual Dense Blocks. Cross-component and cross-view attention blocks, inherited from the baseline architecture, plus text prompts ('albedo' vs 'roughness&metallic') let one network output all material channels consistently across views.","core_discovery":"On its own terms, the paper establishes a recipe for turning a multi-step latent diffusion material estimator into a one-step one without the usual blur. The model is trained at the single timestep t=T, where the latent is pure Gaussian noise and alpha_T=0, so the v-parameterized objective becomes a direct prediction of -z0. Instead of supervising in latent space, the authors decode the prediction and apply pixel-space losses: MSE on albedo and MSE plus a gradient-matching term on the roughness/metallic channels. To recover high-frequency structure lost by the VAE encoder, a Detail Injection Network concatenates early encoder features into late decoder features through Residual Dense Blocks.","pith_inferences":["If the claim that the Gaussian noise input can be removed without hurting performance is correct, a fully deterministic feed-forward material estimator could be distilled from this model, eliminating even the appearance of stochasticity.","The same one-step plus pixel-space loss plus detail-injection recipe is likely portable to other per-pixel regression tasks such as normals, depth, or albedo under unknown lighting, since none of the components is material-specific.","The authors note that DIN sometimes injects small highlights into albedo on glossy textured objects; a natural extension is training DIN with explicit highlight masks or illumination conditioning to separate reflections from texture."],"forward_implications":["Single-view material estimation drops from about 6.1 seconds to about 0.58 seconds and multi-view from about 27.3 seconds to about 2.3 seconds, making diffusion-quality material decomposition practical for batch processing and interactive tools.","Because output variance across noise seeds is low, one sample replaces the ensembles or manual selections that multi-step stochastic estimators currently require.","The Detail Injection Network design transfers to other latent-diffusion dense prediction tasks where VAE encoding blurs contours, text, or other high-frequency structure.","More accurate albedo, roughness and metallic maps without baked reflections feed cleaner priors into inverse-rendering solvers, improving relighting results."],"supporting_citations":[{"why":"Supplies the latent diffusion formulation, the VAE encoder/decoder, and the noise schedule the whole method builds on.","marker":"[Rombach et al. 2022]"},{"why":"IDArb is the main baseline to beat; it also provides the training set and the cross-component/cross-view attention structure reused here.","marker":"[Li et al. 2024]"},{"why":"Shows pixel-space losses make one-step diffusion viable for dense prediction; the paper adopts this for material estimation.","marker":"[Xu et al. 2024]"},{"why":"Provides the v-parameterization used to formulate the one-step training objective.","marker":"[Salimans and Ho 2022]"},{"why":"Supplies the Residual Dense Blocks used to build the Detail Injection Network.","marker":"[Zhang et al. 2018]"},{"why":"Introduces the gradient matching loss that sharpens roughness and metallic boundaries.","marker":"[Li and Snavely 2018]"},{"why":"Provides MIT-Intrinsic, one of the two real-world evaluation sets with pseudo ground truth.","marker":"[Grosse et al. 2009]"},{"why":"ShinyBlender is the synthetic benchmark with ground-truth materials used in the quantitative comparison.","marker":"[Verbin et al. 2022]"},{"why":"Stanford-ORB is the real-world inverse-rendering benchmark used for albedo evaluation.","marker":"[Kuang et al. 2023]"},{"why":"NVDiffRec is the inverse-rendering solver whose decomposition improves when supervised by the model's predictions.","marker":"[Munkberg et al. 2022]"}],"fun_headline_variants":["One-step diffusion model recovers materials with fine detail","Single-step material estimation without blur, low variance","Detail-injected one-step diffusion for sharp material maps","Pixel-space losses keep one-step diffusion materials crisp","One-step diffusion beats multi-step on material accuracy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reported superiority over IDArb rests on a test set the authors built themselves—200 objects selected by excluding IDArb's training indices—so if that selection is unrepresentative of objects or lighting in the wild, the headline gains may shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["One-step diffusion model recovers materials with fine detail","Single-step material estimation without blur, low variance","Detail-injected one-step diffusion for sharp material maps","Pixel-space losses keep one-step diffusion materials crisp","One-step diffusion beats multi-step on material accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1497,"prompt_tokens":743,"completion_tokens":754,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":681}},"tokens_in":487,"tokens_out":754,"duration_ms":7866,"temperature":1.0,"reasoning_tokens":681,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:27:54.924071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an independent set of objects with exact ground-truth albedo, roughness and metallic maps, or the official IDArb test split once released, and compare StableIntrinsic's metrics against IDArb's. If the albedo PSNR gain does not approach 9.9% or the metallic/roughness MSE reductions do not approach 44.4% and 60.0%, the central quantitative claim fails; the same protocol on glossy textured objects tests whether the Detail Injection Network's highlight leakage undermines the detail-preservation claim.","supporting_citations":[],"review_version":1}