{"id":"2c03b81b-dd8d-41a2-ab0a-3074a4b87067","arxiv_id":"2608.08132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ASV3D improves single-view 3D reconstruction by using one extra unposed photo, with a consistency-based gate selecting which image conditions each generated view.","lead":"A new framework, ASV3D, adapts single-view 3D reconstruction with one extra photo of the same object, choosing the best conditioning image for each generated view without needing camera poses. On Google Scanned Objects it reports better geometry and image quality than the original single-view pipelines, and a 32-person user study preferred its outputs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The consistency-based gate in Eq. (4) is never validated as a per-view argmin decision; the central 'extra view helps' claim depends on that decision, yet only pooled correlations on GSO are reported.","rationale":"The reader identifies gate reliability as the weakest assumption. I agree that this is the load-bearing premise, but the sharper problem is internal validation, not only external generalisation. Eq. (4) is an argmin rule, while Appendix A validates only a pooled Spearman correlation between the gate score and generation quality. A positive pooled correlation does not establish that the low-variance condition is the accurate condition in a per-view binary choice; the paper never measures the actual decision accuracy of f. This matters because the optimised variant routes its contrastive loss exclusively through views assigned to z, and the zero-shot variant conditions every view through f. If the gate is wrong on a substantial fraction of views, the method could be suppressing the auxiliary image precisely where it would help. The paper's Table 3 ablation shows the gate outperforms naive concatenation, which is relevant but not decisive. The concrete test proposed here would settle whether the gate's argmin matches the oracle; without that, the SOTA claim rests on an unmeasured decision rule. I do not see grounds to reject the paper outright, because the method is clearly described and the improvement is plausible; however, the conditional acceptance recommended by the reader is appropriate and possibly should explicitly require the gate-decision validation described above.","tokens_in":15686,"tokens_out":4379,"duration_ms":47828,"concrete_test":"Run a held-out oracle-gate comparison. On a leave-one-object-out split of the 30 GSO objects, generate each target view twice with identical seed and trajectory, once conditioned on y and once on z; label the better condition by ground-truth PSNR/LPIPS (and by downstream CD/IoU of the reconstructed mesh) as the oracle choice. Then report: (i) per-view selection accuracy of Eq. (4) against the oracle; (ii) zero-shot reconstruction metrics when f is replaced by oracle, random, and always-y; (iii) the same selection accuracy on the 10 real-world objects using a pseudo-ground-truth such as multi-view consistency or user preference. If selection accuracy is not clearly above chance, or oracle/random/always-y metrics are statistically indistinguishable, the central claim loses its support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The optimised ASV3D-Wonder3D result (CD 0.0218 to 0.0120, IoU 0.5272 to 0.5738) and the zero-shot gains both inherit Eq. (4): every target view is conditioned on f(n), and the contrastive loss in Eq. (10) acts only on views with f(n)={z}. The gate is therefore the mechanism that converts the additional image into a benefit; if f misassigns conditions, the headline numbers do not establish that an extra view helps. Appendix A does not actually test this decision. It reports Spearman correlations between the consistency score g(c) and image-quality metrics over all (condition, view) pairs on GSO, with late-half correlations of +0.390 (negative PSNR) and +0.525 (LPIPS). These are moderate, pooled correlations. A correlation between g and quality does not imply that argmin_c g(c) is the better condition at the per-view level: a consistently low-variance but wrong condition would still win. No held-out objects, no object-level cross-validation, no binary selection accuracy against an oracle, and no real-world gate validation are given. The gate is also computed from the same frozen model that is later optimised, and the optimised variant only uses z on views assigned to z; a biased gate could suppress auxiliary information exactly when it is needed. Table 3 shows the gate beats naive concatenation, but it does not show the gate is correct enough to support the SOTA claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ASV3D, a framework to adapt pre-trained single-view 3D reconstruction models when an additional image of the same object is supplied. Two adaptation strategies are introduced: a zero-shot scheme that uses a consistency-based gate (Eq. 4) to select, per target view, whether the primary or the additional image conditions the diffusion generation, and an optimised scheme that further updates the U-Net with a denoising loss and a contrastive loss (Eqs. 8-11). The method is applied to Wonder3D and Era3D and evaluated on GSO and a small real-world dataset, with quantitative metrics for 3D reconstruction (CD, IoU) and multi-view generation (PSNR, SSIM, LPIPS), plus a user study. The reported results show consistent improvements over the baselines, with the optimised Wonder3D variant achieving the best numbers in both reconstruction and generation.","tokens_in":16029,"tokens_out":5578,"duration_ms":52709,"significance":"If the claimed gating mechanism reliably identifies which input image is more informative for each target view, the paper offers a practical and camera-pose-free way to leverage extra imagery for single-view 3D reconstruction, which is a relevant and timely problem. The zero-shot variant is particularly attractive because it requires no retraining, and the optimised variant demonstrates a promising direction for test-time adaptation. The paper also releases code and a real-world dataset, which are concrete contributions that would benefit the community. However, the central evidence for the gate is currently correlational rather than decision-oriented, and the headline quantitative results inherit this gap; the state-of-the-art claim is therefore not yet fully supported. The core idea is interesting and the experiments are extensive enough that the deficiencies are fixable within the scope of a revision.","major_comments":[{"comment":"The consistency-based gate f chooses, for each target view n, the condition c ∈ {y,z} that minimizes the dispersion of the denoising predictions over the late half of the diffusion trajectory. This gate is the core mechanism that converts the additional image into a reconstruction benefit: it determines the condition used in Eq. (5) for every view, and it also determines which views receive the contrastive loss in Eq. (10). The validation in Appendix A, however, only reports Spearman correlations between the consistency score g(c) and image-quality metrics pooled over all (condition, view) pairs on GSO (Figure 8). A pooled correlation does not establish that argmin_c g(c) selects the better condition for a specific view; a condition that consistently yields low dispersion but poor fidelity would win the argmin despite being the wrong choice. The paper does not report per-view binary selection accuracy against an oracle, object-level cross-validation, or any test of the gate on held-out objects or real-world data. Table 3 shows that the gate beats a multi-conditioning baseline, but that comparison does not separate the gate's decision accuracy from the general benefit of selective conditioning. Consequently, the zero-shot improvements in Table 1 and the condition assignments used in the optimised variant are not convincingly shown to arise from the gate correctly identifying which input image is more informative for each target view, which is the central claim of the paper.","section":"Sec. 3.1, Eq. (4) and Appendix A"},{"comment":"The hyperparameters of the method—the contrastive weight λ (set to 0.2), the temperature τ (set to 0.07), the number of diffusion steps T (set to 50), and the late-half step range T in Eq. (4)—are all selected on the same GSO benchmark that is used for the main quantitative evaluation. No cross-validation on held-out objects or a separate validation set is reported, and no sensitivity analysis is given for τ, T, or the late-half range. Because the paper frames the method as a test-time adaptation scheme, the absence of such analysis leaves open the possibility that the reported gains rely on hyperparameters that have been tuned to the evaluation distribution. At least a sensitivity table or a statement about the stability of the results under reasonable hyperparameter variations is needed for the claims in Table 1 to be fully load-bearing.","section":"Sec. 4.2 and Sec. 4.4"},{"comment":"The user study is reported with mean ratings and standard deviations, but no significance testing is performed, and the standard-deviation values in Figure 7 appear to contradict the claim in the text that ASV3D has 'lower standard deviations' than Wonder3D: the figure shows ASV3D multi-view generation with ±0.53 and Wonder3D with ±0.12, which would imply the opposite. Please clarify the figure or correct the text, and add appropriate statistical tests (e.g., paired t-test or Wilcoxon signed-rank) on the ratings or the forced-choice preferences.","section":"Sec. 4.5 and Figure 7"}],"minor_comments":[{"comment":"The tables report point estimates without error bars or significance tests; given that only 30 GSO objects are used, per-object standard deviations or confidence intervals would help assess the stability of the differences.","section":"Tables 1, 3, 4"},{"comment":"There is a recurring typo: 'FreeSplater' should be 'FreeSplatter'.","section":"Figures 4 and 10"},{"comment":"The description of how the additional image is rendered for GSO (\"another image in a random view\") is underspecified; please state whether the view is uniformly sampled, whether it is always a different azimuth, and whether any objects are excluded due to near-duplicate views.","section":"Sec. 4.1"},{"comment":"The symbol T is used both for the total number of diffusion steps and for the set of late-half step indices; using a different symbol for the set (e.g., T_late) would avoid confusion.","section":"Eq. (4)"},{"comment":"The real-world evaluation is qualitative only (10 objects), which is understandable because ground-truth 3D is unavailable; however, this limitation should be stated more prominently in the main text rather than only in the supplementary.","section":"Real-world evaluation"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the validation of the gating mechanism, which is the load-bearing contribution of the paper. I recommend the revision add an oracle comparison on GSO (for each target view, generate with both conditions and label the better result per a metric; then report gate accuracy), held-out-object experiments, and at least a qualitative gate check on real-world data. The user-study standard deviation inconsistency and the lack of statistical testing should also be addressed. These are fixable within the manuscript's scope, so I recommend major revision rather than reject. If the oracle experiments reveal that the gate is not reliable, the paper would need to weaken its central claim substantially."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is the per-view consistency gate: given a primary image and one extra unposed image, it picks, for each target view, which image should condition the diffusion model. That is a new idea, and the paper sells it clearly. The optimised variant adds a contrastive loss on views assigned to the auxiliary image, and the reported gains on Wonder3D are substantial (Chamfer 0.0218 to 0.0120 on GSO). I also note that they promise code and a real-world dataset, which should count in their favor.\n\nWhat the paper does well: the method is simple to state, the derivations are standard, and the experimental protocol is mostly transparent. The ablation in Table 3 is useful—it shows that per-view selection beats naively concatenating both images, which supports the central claim that the extra view helps when routed properly. The user study is a plus, even if small.\n\nThe soft spots are real but not fatal. The gate is the load-bearing mechanism, yet Appendix A validates it only via pooled Spearman correlations between the gate score and image quality over all (condition, view) pairs on GSO. That does not directly test the actual decision rule: whether argmin_c g(c) picks the better condition for a given view. A moderate correlation can hold while a consistently low-variance wrong condition still wins. No per-view binary accuracy, no held-out objects, no real-world gate evaluation. The main tables also lack error bars, hyperparameters (lambda, late-half range) are tuned on the same benchmark, and real-world results are qualitative only. These limitations cap the strength of the SOTA claim, but they do not contradict the core improvement.\n\nWho should read this: researchers working on test-time adaptation or conditioning strategies for generative 3D reconstruction. It is an incremental but useful contribution, not a paradigm shift.\n\nMy recommendation: send it to serious peer review. The idea is worth engaging, the code will help, and the requested revisions are concrete: error bars, a direct evaluation of the gate's per-view selection accuracy (oracle comparison), and ideally validation on a held-out set. If those land, the paper would be solid.","headline":"A genuinely new per-view conditioning gate for adding an unposed extra image to single-view 3D reconstruction, with a real but narrow validation gap around gate correctness.","tokens_in":16536,"tokens_out":2875,"would_cite":false,"duration_ms":30441,"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":"ASV3D lets one extra image improve single-view 3D reconstruction by choosing the best condition for each generated view.","keywords":["test-time adaptation","single-view 3D reconstruction","multi-view generation","consistency-based gating","contrastive learning","diffusion models","zero-shot adaptation","additional image"],"falsifier":"Take objects from a category outside GSO, such as glassware or symmetric toys, and for each target view compute both the gate's chosen condition and the actual reconstruction error under both conditions. If the gate's selections do not correlate with the lower-error condition, or if the Spearman correlation between gate consistency and generation quality drops below its GSO values, the proxy collapses.","tokens_in":15506,"feed_emoji":"📷","tokens_out":5853,"duration_ms":48570,"temperature":0.7,"pith_summary":"ASV3D is a framework that adapts a pre-trained single-view 3D reconstruction model to test time by supplying one additional image of the same object, captured by a different camera and in a different context. It addresses the question of when an extra view helps: rather than fusing all images into one condition, ASV3D uses a consistency-based gate to choose, for each target view, whichever conditioning image the diffusion model denoises most consistently. The zero-shot variant does this without retraining; the optimised variant further fine-tunes the U-Net's attention layers with a denoising loss and a contrastive loss that pulls views generated from the additional image closer to that image. The authors report consistent gains over Wonder3D and Era3D on the GSO benchmark, and on a small real-world dataset, with the optimised Wonder3D reducing Chamfer distance from 0.0218 to 0.0120. The aim is to show that a single second image can resolve some of the ambiguity of single-view reconstruction without pose estimation or retraining from scratch.","feed_headline":"Extra image sharpens single-view 3D reconstruction","feed_subtitle":"Consistency-based gating picks the more informative input per view, cutting Chamfer distance by 45 percent.","key_machinery":"The consistency-based gate f(n) defined in Eq. (4) is the load-bearing mechanism: for each target view index n, it picks the condition c in {y,z} that minimises the expected squared deviation of the denoising-mean prediction mu from its own time-average over the late half of the diffusion trajectory, T = {T/2, ..., 1}. This is a pose-free proxy for which input image is most informative for that view, replacing the need to estimate camera poses or fuse conditions. In the optimised variant, a contrastive loss (NT-Xent, Eq. 9-10) anchored on the additional image z is added to the denoising loss, and only the cross- and self-attention layers of the U-Net are updated.","core_discovery":"The paper's central claim is that a single additional image of the same object, even from an unposed, different camera, can be folded into an existing single-view generative reconstruction pipeline to improve both multi-view generation and the final 3D shape. The mechanism is a camera-pose-free consistency gate: for each target view n, it evaluates the variance of the diffusion model's denoising-mean predictions µ over the late half of the denoising trajectory under each candidate condition, and selects the condition with the smaller spread. Under the optimised strategy, the model is then adapted by minimising a denoising loss plus a contrastive (NT-Xent) loss that anchors views generated with the additional image to that image. On Google Scanned Objects, the optimised Wonder3D variant makes Chamfer distance 0.0120, IoU 0.5738, and improves PSNR/SSIM/LPIPS over the baseline; a 32-participant user study rates ASV3D higher on both 3D reconstruction and multi-view generation. The authors position this as the first method to adapt single-view reconstruction with additional imagery.","pith_inferences":["We infer that the gate's Spearman correlation with generation quality, verified only on GSO, may not hold for objects with strong symmetries or specular surfaces, where small denoising variance could arise from a confident but wrong condition; testing the gate on such categories would be a direct stress test.","Beyond the paper's explicit claim, the method suggests a simple acquisition policy: when generating a target view, the gate could rank candidate supplementary images by their consistency scores and pick the best one from a pool, turning the binary choice into a selection among many views.","We also conjecture that the contrastive loss could be replaced or augmented by attention-based feature alignment between the generated views and the additional image, which might reduce the need for a separate image encoder; this is an editorial suggestion, not a claim of the paper.","The reported real-world evaluation is qualitative, so a natural next step is a quantitative evaluation with ground-truth scans of objects in natural settings, which the paper does not provide."],"forward_implications":["Any existing conditional generative single-view reconstruction model can likely benefit from the same two-stage adaptation, since the gate only uses the model's own denoising predictions.","Because the gate decides per target view, the framework naturally handles cases where the extra view is helpful for some views (e.g., occluded back views) and redundant or misleading for others (e.g., views close to the primary image).","The optimised variant's reliance on contrastive learning to pull z-conditioned views toward z should transfer to improving cross-view consistency in other multi-view diffusion pipelines.","If the gate correctly identifies informative conditions, the method could be used to tell a user which additional view to acquire for a given object, a direction the authors mention as future work.","The improvements are obtained without pose estimation or calibration, opening test-time adaptation for casual multi-view inputs that are both unposed and captured under different conditions."],"supporting_citations":[{"why":"Wonder3D, the primary single-view baseline that ASV3D adapts; supplies the U-Net and cross-domain diffusion setup.","marker":"(Long et al. 2024)"},{"why":"Era3D, the second baseline used to show the generality of the adaptation.","marker":"(Li et al. 2024)"},{"why":"Defines the DDPM forward and reverse equations that the generation process in Eqs. (1)-(3) is built on.","marker":"(Ho, Jain, and Abbeel 2020)"},{"why":"Consistency models, the conceptual basis for the consistency-based gate's reliance on denoising convergence.","marker":"(Song et al. 2023)"},{"why":"Consistent diffusion models, also cited as inspiration for measuring denoising consistency in the gate.","marker":"(Daras et al. 2023)"},{"why":"NT-Xent loss, used verbatim in Eq. (9) for the contrastive term of the optimised adaptation.","marker":"(Chen et al. 2020)"},{"why":"Google Scanned Objects dataset used for all quantitative comparisons and gate validation.","marker":"(Downs et al. 2022)"},{"why":"NeuS, used to reconstruct the 3D model X from the generated views x0^1:N.","marker":"(Wang et al. 2021)"}],"fun_headline_variants":["Extra view cuts 3D reconstruction error by 45%","One extra image adapts single-view 3D without retraining","Test-time extra photo improves 3D reconstruction accuracy","ASV3D: extra view sharpens single-view 3D reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the consistency-based gate of Eq. (4), which equates the variance of the model's denoising-mean predictions over the late half of the trajectory with the reliability of a conditioning image, is a valid proxy for which input produces the more accurate target view.","fun_headline_variants_meta":{"raw":{"variants":["Extra view cuts 3D reconstruction error by 45%","One extra image adapts single-view 3D without retraining","Test-time extra photo improves 3D reconstruction accuracy","ASV3D: extra view sharpens single-view 3D reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1644,"prompt_tokens":1005,"completion_tokens":639,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":566}},"tokens_in":621,"tokens_out":639,"duration_ms":6256,"temperature":1.0,"reasoning_tokens":566,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:22:09.254941+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take objects from a category outside GSO, such as glassware or symmetric toys, and for each target view compute both the gate's chosen condition and the actual reconstruction error under both conditions. If the gate's selections do not correlate with the lower-error condition, or if the Spearman correlation between gate consistency and generation quality drops below its GSO values, the proxy collapses.","supporting_citations":[],"review_version":1}