{"id":"fb3424c4-7ebc-4bad-8c0f-6cd6f91c5c34","arxiv_id":"2411.18866","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RIGI improves image-to-3D generation by estimating pixel-wise uncertainty from the difference between two 3D Gaussian models and using it to reweight the reconstruction loss, reducing artifacts from inconsistent multi-view pseudo-labels.","lead":"This paper adds uncertainty-aware learning to 3D Gaussian Splatting for image-to-3D generation, using the disagreement between two concurrently trained 3D models to down-weight conflicting multi-view labels and reduce artifacts. It reports modest quantitative gains over state-of-the-art baselines and stronger user-study preference.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncertainty map may encode generic model variance rather than pseudo-label inconsistency; loss can act as a smoother, leaving central mechanism unverified.","rationale":"After re-reading the paper, I agree with the reader's CONDITIONAL verdict and with their identification of the weakest assumption: nowhere is U shown to track pseudo-label errors. I examined Eq. (5)-(8) and Section 3.2-3.3, and the internal discussion in Section 4.2 (Figure 7) reinforces the concern by showing that aggressive λ produces exactly the over-smoothing that a generic agreement regularizer would cause. The paper has strengths: the two-model ensemble is simple, efficient, and the qualitative results are compelling; the use of PAG and dynamic elevation sampling are reasonable engineering contributions. However, the quantitative evaluation is thin (25 objects, no error bars, no code), and no ablation isolates the uncertainty term from the other components. The central mechanism could still be correct, but the paper as written does not rule out the alternative explanation that the benefit comes from a generic smoothing/ensembling effect. The proposed oracle-correlation test would settle this. Since the concern is addressable with additional experiments and does not disprove the method, I leave the verdict as CONDITIONAL.","tokens_in":13459,"tokens_out":4785,"duration_ms":46408,"concrete_test":"On a fixed set of 10 GSO objects, generate three independent SV3D videos for the same 63-pose camera trajectory, so each viewpoint has three pseudo-label samples. Train two 3DGS models with the standard reconstruction loss (no uncertainty weighting), and at iteration 5000 render U = |I1 - I2| at all poses. Compute per-pixel variance across the three pseudo-label samples as an oracle inconsistency map, and report the Spearman rank correlation between U and this oracle map. If the correlation is weak (<0.3), the premise that U proxies pseudo-label inconsistency is not supported. Additionally, rerun the uncertainty-aware training with U replaced by a fixed random map; if the artifact reduction and metric changes are statistically indistinguishable from the proposed method, the mechanism is generic smoothing rather than targeted rectification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The causal chain of the paper hinges on Eq. (5): U = |I1 - I2| is taken to indicate where the pseudo-labels from SV3D are inconsistent, so that reweighting the loss in Eq. (6) can rectify those conflicts. The manuscript provides no direct evidence for this correlation. Two independently initialized 3DGS models optimized on the same frames will also diverge because of sparse-view ambiguity, densification randomness, local minima, and optimization instability; this is generic epistemic variance, not necessarily aleatoric label noise. Moreover, the second term λU in Eq. (6) penalizes disagreement between the two models directly, forcing I1 and I2 toward each other everywhere. High-U regions are therefore suppressed regardless of whether the label there is wrong, and the method will smooth the result even if every pseudo-label is perfectly consistent. The authors' own λ=10 ablation (Fig. 7) shows exactly this failure mode. Because the full system also adds PAG, dynamic elevation sampling, progressive sampling, and random background colors, with no quantitative ablation isolating the uncertainty term, the qualitative gains do not establish that inconsistency is being detected and mitigated; they are also consistent with the hypothesis that the method is a generic low-pass regularizer that trades fidelity for smoothness (consistent with the lower PSNR vs. Hi3D in Table 1). This is the weakest load-bearing point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes RIGI, an image-to-3D generation method that combines SV3D-generated multi-view pseudo-labels with 3D Gaussian Splatting optimization. The central idea is to optimize two randomly initialized Gaussian models G1 and G2 simultaneously, define an uncertainty map as the absolute difference between their rendered images from the same viewpoint (Eq. 5), and use this map to reweight the reconstruction loss (Eq. 6), down-weighting supervision in high-uncertainty regions. The full pipeline also includes Perturbed-Attention Guidance, dynamic elevation sampling, progressive sampling, and random background colors. The experimental section reports results on 25 GSO objects with PSNR/SSIM/LPIPS, a user study, and qualitative ablations, and the paper claims that the method reduces artifacts and floats caused by inconsistent pseudo-labels while improving structural and perceptual quality.","tokens_in":13699,"tokens_out":7068,"duration_ms":60611,"significance":"If the central premise is valid, the dual-model disagreement map is a simple and training-stable way to down-weight conflicting pseudo-labels without learning an auxiliary variance head, and it could be integrated into many reconstruction-based image-to-3D pipelines. The paper has several strengths: it evaluates against external GSO ground truth rather than only the pseudo-labels, it includes a user study, and it provides qualitative ablations of the uncertainty weight and of alternative uncertainty designs. However, the current evidence does not yet establish that the uncertainty map localizes label inconsistency rather than generic model variance, and the reported quantitative gains are small and unaccompanied by error bars. The idea is interesting and testable, but the manuscript needs substantially stronger experiments before the central claim is supported.","major_comments":[{"comment":"The central premise that U = |Î1 - Î2| localizes inconsistency in the generated pseudo-labels is not validated. Two randomly initialized 3DGS models optimized on the same frames will also diverge because of sparse-view ambiguity, densification randomness, and optimization instability, so the disagreement map can capture generic epistemic variance rather than label conflict. Furthermore, the second term +λU in Eq. (6) penalizes disagreement between G1 and G2 in every pixel, which suppresses high-U regions even when the pseudo-label is correct; this makes the loss act as a low-pass regularizer. The λ=10 result in Fig. 7 is consistent with this concern, as it produces smoother and blurrier outputs. The authors should provide direct evidence of a correlation between U and actual pseudo-label errors, for example by injecting known inconsistencies into otherwise consistent labels and measuring whether U identifies them, or by comparing U against per-pixel error maps in SV3D frames.","section":"3.2, Eq. (5)"},{"comment":"The reconstruction loss includes an uncertainty-weighted L1 term but also an unweighted LPIPS term, Llpips = LPIPS(Igt, Î1) + LPIPS(Igt, Î2), with weight λ_l = 0.5. Because LPIPS is applied at full strength in all pixels, high-uncertainty regions still receive strong supervision from potentially inconsistent pseudo-labels, which contradicts the stated goal of reducing reconstruction intensity in those regions. The authors should either remove or uncertainty-weight the LPIPS term in an ablation, or explain why LPIPS is exempt from the uncertainty reweighting.","section":"3.3, Eqs. (7)-(8)"},{"comment":"The quantitative evidence is too thin to support the headline claim. The evaluation uses only 25 GSO objects, reports no standard deviations or significance tests, and the proposed method does not achieve the best PSNR (Hi3D reaches 17.2559 vs. 16.9646 for RIGI). The reported gains are small: SSIM improves by 0.0129 over Hi3D and LPIPS by 0.0011. With no error bars, 'superior or comparable' is not established. In addition, the pipeline comparison is uncontrolled: the baselines do not share the same SV3D+PAG frame generation, elevation sampling, progressive sampling, or random-background tricks, so any metric difference could come from those components rather than from uncertainty-aware learning. Please report per-object scores, confidence intervals, and a controlled comparison where only Eq. (6) is toggled.","section":"4.1, Table 1"},{"comment":"The ablations do not isolate the uncertainty term. Figure 5 adds progressive sampling and uncertainty-aware learning simultaneously; Figure 6 compares uncertainty designs only qualitatively; and Figure 7 varies λ without reporting quantitative metrics. A quantitative ablation in the full pipeline, with and without the uncertainty term and with all other components fixed, is necessary to demonstrate that the method detects and rectifies inconsistency rather than merely smoothing the output. Ideally the authors would also report a metric that directly measures label inconsistency or artifacts, such as the number of floaters or edge error.","section":"4.2, Figures 5-7"}],"minor_comments":[{"comment":"There is a typo in the sentence 'The first term dynamically adjusts the optimation intensity'; 'optimation' should be 'optimization'.","section":"3.3"},{"comment":"The text says 'we progressively increase the render radio'; this should be 'render ratio' or 'rendering resolution'.","section":"3.4"},{"comment":"The symbol λ is used for the D-SSIM weight in Section 3.1 (λ=0.2) and for the uncertainty weight in Eq. (6) (λ=5); renaming one of these would avoid notational collision.","section":"3.1 and 3.3"},{"comment":"Because each participant selects two outputs, the preference percentages sum to 200%; this should be stated in the caption or table, and confidence intervals should be reported.","section":"Table 2"},{"comment":"The caption has a duplicated auxiliary verb: 'may make the generated results may become smoother and blurrier'.","section":"Figure 7"},{"comment":"The caption contains a spacing error, 'wP A G', which should read 'w/ PAG'.","section":"Figure 8"},{"comment":"The relationship between the 25 GSO objects used for metric evaluation and the 30 samples used in the user study should be clarified.","section":"4, Evaluation Metrics vs. User Study"},{"comment":"Reference [52] is typeset as 'InECCV' without a space; the formatting should be corrected.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea is testable, so I am not recommending rejection. The main gap is experimental: the central mechanism linking the disagreement map to pseudo-label inconsistency is never directly verified, and the ablations are qualitative and confounded with other pipeline components. If the authors can provide a controlled quantitative ablation and a correlation analysis between U and label errors, the paper would be much stronger. I would also encourage the editor to ask for code and random-seed release, given the small reported metric differences."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's new bit is simple: take two 3D Gaussian models optimized on the same SV3D pseudo-frames, use the absolute rendered-image difference as a pixel-wise uncertainty map, and down-weight the reconstruction loss where that map is high. Applied to this specific pipeline stage, that is a legitimate new use of a known ensemble heuristic. The qualitative results are believable—the method does seem to clean up edge artifacts and floats—and the ablations cover the obvious design alternatives (learnable variance vs. two-model ensemble). The authors are also honest about trailing Hi3D on PSNR and offer a plausible explanation.\n\nThe soft spot is the one the stress-test note names: the disagreement map is never shown to correlate with actual pseudo-label errors. Two randomly initialized Gaussians will diverge from sparse-view ambiguity, densification randomness, and optimization instability, not only from label conflict. The second loss term, λU, actively pulls the two models together everywhere, so the method can smooth even when every pseudo-label is consistent. The paper's own λ=10 ablation (Figure 7) shows exactly that blurring failure, which supports the concern. A correlation analysis against known inconsistencies (or a synthetic test with injected label noise) would have tied the mechanism down. Without it, the observed improvements could be explained by generic regularization rather than uncertainty-guided rectification.\n\nOther weaknesses are proportionate: 25 objects, no error bars, no code, several tuned hyperparameters (λ, λs, λl, elevation schedule), and no quantitative ablation that isolates the uncertainty term from PAG, dynamic sampling, and progressive rendering. These are addressable, not fatal. The contribution is incremental—disagreement-based uncertainty is established—so the novelty is in the application, not the principle.\n\nOverall: a clear, honest paper with a plausible idea and believable qualitative gains, but the load-bearing claim about inconsistency detection remains unverified. It deserves a serious referee, who should push for stronger evidence on what the uncertainty map actually encodes. I would not cite it in my own work until that is clarified, but it is a good reading-group discussion candidate.","headline":"A plausible uncertainty-weighted regularization trick for image-to-3D, but the central mechanism is not directly validated and the evaluation is thin.","tokens_in":14247,"tokens_out":2042,"would_cite":false,"duration_ms":20759,"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":"The paper claims that pixel-wise disagreement between two concurrently optimized 3D Gaussian models estimates where generated multi-view labels conflict, and reweighting the reconstruction loss by this disagreement reduces artifacts and…","keywords":["image-to-3d generation","3d gaussian splatting","uncertainty-aware learning","multi-view consistency","pseudo-label rectification","single-image 3d reconstruction","uncertainty regularization"],"falsifier":"Render a synthetic object with a known, localized inconsistency, such as an incorrect texture patch pasted onto one of the multi-view labels, then check whether the uncertainty map $U$ shows a spike exactly at that patch and whether suppressing it restores the clean geometry.","tokens_in":13234,"feed_emoji":"🎲","tokens_out":6178,"duration_ms":53395,"temperature":0.7,"pith_summary":"This paper tries to solve a common failure mode in image-to-3D generation: the multi-view frames that guide 3D reconstruction are synthesized by a diffusion model, and those frames often contradict one another in geometry or texture, producing boundary artifacts and floating debris in the final 3D asset. The authors' proposal is to run two copies of the 3D Gaussian Splatting reconstruction at the same time, render both from the same camera, and treat the absolute difference between the two renders as an uncertainty map. Where the two models disagree, pseudo-label supervision is down-weighted; where they agree, full supervision is kept. The claimed effect is that the optimizer stops over-reconstructing inconsistent regions, yielding smoother geometry and cleaner edges, with structural and perceptual metrics improving even though pixel-level PSNR stays roughly the same.","feed_headline":"Twin 3D models flag conflicting views to fix 3D generation artifacts","feed_subtitle":"Two co-trained Gaussian models detect where synthetic labels disagree, reweighting the loss to reduce boundary artifacts and floats.","key_machinery":"The machinery is the twin-model disagreement estimator. Two randomly initialized 3D Gaussian Splatting models are trained in parallel on the same multi-view pseudo-labels; at each optimization step, both are rendered from the same sampled viewpoint, and the absolute pixel difference $U = |\\hat{I}_1 - \\hat{I}_2|$ becomes the uncertainty signal. That map is folded into an uncertainty-weighted L1 loss, $L_u = |I_{\\text{gt}} - \\hat{I}| / \\exp(\\lambda U) + \\lambda U$, so high-disagreement regions are regularized instead of aggressively reconstructed. The second term, $\\lambda U$, acts as a stabilizer that prevents the two models from diverging globally. The mechanism runs inside standard 3DGS training and requires only two forward renders rather than a learnable uncertainty branch.","core_discovery":"The central claim is that uncertainty is observable as disagreement between two stochastic reconstructions, and that routing this uncertainty into the loss rectifies the reconstruction. Concretely, two 3D Gaussian models $G_1$ and $G_2$ are optimized simultaneously on the same set of pseudo-label frames; for a sampled camera pose, both are rendered, and the uncertainty map is $U = |\\hat{I}_1 - \\hat{I}_2|$. The per-pixel loss becomes $|I_{\\text{gt}} - \\hat{I}| / \\exp(\\lambda U) + \\lambda U$, with $\\lambda = 5$, so pixels with strong model disagreement are sharply down-weighted while the additive $\\lambda U$ term keeps the two models from drifting too far apart. Over optimization this adjusts the effective supervision per pixel, reducing the pressure to densify Gaussian points in regions where the generated frames conflict. The paper reports that this suppresses artifacts and floats, improves SSIM and LPIPS, and keeps PSNR nearly unchanged, interpreting that pattern as a shift from exact pixel matching toward structural coherence in inconsistent areas.","pith_inferences":["A natural next test is whether the disagreement map predicts actual ground-truth label error rather than generic model variance, by corrupting known pixels in a controlled pseudo-label set and measuring the precision of $U$.","The twin-model trick could transfer to feed-forward reconstruction models as a soft-label weighting scheme during training, or as a confidence prior for choosing which generated frames to trust.","The uncertainty map could also serve as an automatic cleanup signal after generation: regions flagged as consistently uncertain across many frames could be re-rendered or re-sampled rather than simply down-weighted.","Because $U$ is computed purely from rendered images, the same procedure could be applied to multi-view frames produced by any generative model, regardless of the specific video diffusion backbone."],"forward_implications":["Two-stage image-to-3D pipelines can suppress pseudo-label conflicts at optimization time without retraining or modifying the diffusion model that produces the frames.","The uncertainty map gives a per-pixel confidence signal that can be logged during optimization to locate where synthetic frames disagree, turning an invisible failure mode into a measurable quantity.","Because the weighting is pixel-wise and model-agnostic, the same loss reweighting can be applied alongside other 3D Gaussian reconstruction losses, not only the L1 term used in the paper.","Under the paper's assumption, artifacts such as boundary floats and extra geometric protrusions should shrink in high-uncertainty regions, matching the qualitative ablation evidence the paper reports."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation and base reconstruction loss that the uncertainty weighting modifies.","marker":"[19]"},{"why":"The multi-view video diffusion model that produces the dense pseudo-label frames; their inconsistency is the problem being addressed.","marker":"[55]"},{"why":"Perturbed-attention guidance integrated into the video diffusion model to improve the structural quality of generated frames.","marker":"[1]"},{"why":"LPIPS perceptual loss included in the total reconstruction loss alongside the uncertainty-weighted L1 term.","marker":"[68]"},{"why":"D-SSIM structural loss used as part of the reconstruction objective that the uncertainty weighting sits alongside.","marker":"[59]"}],"fun_headline_variants":["Twin 3D models spot disagreement to fix artifacts","Uncertainty-aware loss cleans up 3D generation noise","Disagreement maps repair inconsistent 3D views","Two Gaussian models dampen high-risk pixels in 3D","Uncertainty-weighted training reduces 3D boundary glitches"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that when two copies of the same 3D model, trained from different random starts on the same synthetic frames, disagree at a pixel, the reason is that the synthetic labels conflict there, not merely that the model is uncertain for unrelated reasons.","fun_headline_variants_meta":{"raw":{"variants":["Twin 3D models spot disagreement to fix artifacts","Uncertainty-aware loss cleans up 3D generation noise","Disagreement maps repair inconsistent 3D views","Two Gaussian models dampen high-risk pixels in 3D","Uncertainty-weighted training reduces 3D boundary glitches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1380,"prompt_tokens":1004,"completion_tokens":376,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":293}},"tokens_in":620,"tokens_out":376,"duration_ms":4525,"temperature":1.0,"reasoning_tokens":293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:48:00.087060+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a synthetic object with a known, localized inconsistency, such as an incorrect texture patch pasted onto one of the multi-view labels, then check whether the uncertainty map $U$ shows a spike exactly at that patch and whether suppressing it restores the clean geometry.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting representation and base reconstruction loss that the uncertainty weighting modifies."},{"cited_title":"Sv3d: Novel multi-view syn- thesis and 3d generation from a single image using latent video diffusion","cited_arxiv_id":null,"evidence_quote":"The multi-view video diffusion model that produces the dense pseudo-label frames; their inconsistency is the problem being addressed."},{"cited_title":"Image quality assessment: from error visibility to structural similarity","cited_arxiv_id":null,"evidence_quote":"D-SSIM structural loss used as part of the reconstruction objective that the uncertainty weighting sits alongside."}],"review_version":1}