{"id":"d593aa40-3867-4b5d-b952-1d6685527b01","arxiv_id":"2506.21034","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"DidSee is a diffusion-based depth completion model that combines a zero terminal-SNR noise scheduler, single-step training, and a semantic segmentation enhancer to achieve state-of-the-art results on non-Lambertian objects.","lead":"A diffusion-based method, DidSee, restores missing depth for transparent and shiny objects by fine-tuning Stable Diffusion with a single-step noise-agnostic training scheme and a semantic enhancer. DidSee reports state-of-the-art results on transparent-object depth benchmarks and improves robotic grasping and pose estimation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The single-step-from-noise formulation assumes the U-Net learns a stable deterministic mapping from pure Gaussian noise to depth; without a noise-invariance analysis, the reported SoTA gains could be seed-dependent.","rationale":"The reader's weakest assumption is on the single-step-from-noise formulation, and I agree this is the most load-bearing point. The central claim is that DidSee achieves SoTA performance and robust real-world generalization. For that to be true, the model must produce accurate depth latents in a single forward pass from a pure-noise latent. This is a much harder conditional generation problem than iterative denoising, and the paper's evidence consists entirely of point estimates. The absence of error bars or multi-seed analysis is especially relevant because the noise input is unconstrained; if different noise samples produce different depth maps, the reported numbers are not well-defined. The ablation confound (single-step plus new loss) also means the paper does not isolate the proposed mechanism. A noise-invariance test would settle whether the concern lands. If the variance is small, the central claim is substantially supported; if large, the method is not noise-agnostic and the reported gains are not reproducible. I therefore keep the verdict unchanged: the empirical results may be correct, but they require this verification.","tokens_in":15810,"tokens_out":9663,"duration_ms":123547,"concrete_test":"Reproduce or obtain the trained DidSee model and run inference on the TransCG validation set 20 times, each time sampling a fresh Gaussian noise vector eps for the t=T latent while fixing RGB, raw depth, and all other settings. Compute the mean and standard deviation of RMSE across the 20 runs. If the standard deviation exceeds about 5% of the mean, or if the best seed outperforms the median by more than the 15% RMSE margin reported over TODE-Trans, the result is seed-dependent and the noise-agnostic assumption fails. As a secondary check, retrain the single-step model with the original latent-space v-prediction loss instead of the pixel-space L1 loss; if performance is unchanged, the improvement attributed to single-step training is actually due to the task-specific loss.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Eq. (7), zero terminal-SNR sets alpha_T=0, so the v-prediction target becomes v_T = z0, independent of the sampled noise. But the model input at t=T is zy_T = eps, a fresh standard Gaussian latent. The network must therefore learn to be invariant to an uninformative, randomly sampled input channel. Nothing in the U-Net architecture enforces this invariance; the expected-loss minimizer is the conditional median, but a finite-capacity network trained with SGD can retain substantial sensitivity to eps. The paper does not report output variance across noise samples, does not specify the inference-time noise seed, and provides no code. Consequently, the single-number SoTA results in Tables 2-5 could be seed-dependent, which directly threatens the 'robust generalization' claim. The ablation in Table 6 compounds this: the move to single-step training is introduced together with the pixel-space task-specific loss (Sec. 3.4), so the REL drop from 0.020 to 0.017 cannot be attributed solely to noise-agnostic training. If the observed gains come from the loss rather than the single-step formulation, or if the model's output fluctuates with eps, the central claim that the proposed diffusion modifications are responsible for SoTA performance is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DidSee, a Stable-Diffusion-based depth completion method for non-Lambertian (transparent/specular) objects. It identifies two biases in vanilla diffusion: signal leakage from a nonzero terminal SNR and exposure bias from multi-step sampling. It addresses them with a rescaled zero-terminal-SNR scheduler, a noise-agnostic single-step training formulation with a pixel-space L1 task loss, and a semantic enhancer that jointly predicts depth and palette-coded semantic maps via task switchers and cross-task attention. Experiments on TransCG, ClearPose, and DREDS/STD report state-of-the-art depth metrics, ablations for each component, and improved downstream category-level pose estimation and robotic grasping.","tokens_in":16015,"tokens_out":8441,"duration_ms":91803,"significance":"If the reported results hold, DidSee is a meaningful step: it shows that two diffusion-specific biases are material for depth completion and that a single-step zero-SNR formulation with a task-specific loss plus semantic conditioning can outperform task-specific transparent-object depth completion methods on three benchmarks and improve downstream robotics. The paper's strengths are the clear decomposition of the design (zero-SNR scheduler, single-step training, semantic enhancer), comprehensive comparison against strong baselines, and inclusion of downstream evaluations. However, the load-bearing issues below—the v-prediction reconstruction identity, the unexamined dependence on the random input noise at t=T, and the confounded ablation—must be resolved before the state-of-the-art claim is fully established; the semantic enhancer also lacks any semantic-level evaluation and training-data specification. No code is provided, so the reported single-number results cannot be independently reproduced at this stage.","major_comments":[{"comment":"The reconstruction formula in Eq. (4) is inconsistent with the v-prediction target in Eq. (2). With v_t = sqrt(alpha_t) eps + sqrt(1-alpha_t) z0 and z_t = sqrt(alpha_t) z0 + sqrt(1-alpha_t) eps, substituting into hat_z0 = sqrt(alpha_t) z_t + sqrt(1-alpha_t) hat_v_t gives hat_z0 = z0 + 2 sqrt(alpha_t(1-alpha_t)) eps, not z0. The identity is exact only at the endpoints alpha_t=0 or 1, so the single-step t=T case is unaffected, but the multi-step baseline, Fig. 4, and the Table 6 baseline rely on this formula for intermediate timesteps. The authors should correct the sign convention (standard v-prediction uses v_t = sqrt(alpha_t) eps - sqrt(1-alpha_t) z0 and hat_z0 = sqrt(alpha_t) z_t - sqrt(1-alpha_t) hat_v_t) and recompute the affected experiments.","section":"Sec. 3, Eqs. (2)-(4)"},{"comment":"The 'noise-agnostic' claim is not established for a finite-capacity network. Setting alpha_T=0 makes the training target v_T = z0 independent of eps, but the model input is z_T = eps, a fresh standard Gaussian sample. The optimal predictor is the conditional mean, but nothing in the U-Net architecture forces invariance to this uninformative input, and SGD may leave substantial sensitivity. The paper does not report the variance of the restored depth over noise realizations at inference, does not give the noise seed used for Tables 2-5, and reports all metrics without error bars. Please add a noise-invariance experiment (e.g., mean +/- std of RMSE/REL over several eps at fixed inputs) and report seeds; without this, the single-number SoTA results could be seed-dependent.","section":"Sec. 3.2, Eq. (7)"},{"comment":"The '+Noise-Agnostic Single-step Training' ablation changes two variables simultaneously: the sampling schedule is moved to one step at t=T and the training objective is switched from the standard diffusion/latent objective to the pixel-space task-specific loss in Eq. (8). The REL drop from 0.020 to 0.017 therefore cannot be attributed solely to the single-step formulation, and the subsequent claim that this row demonstrates the adverse effect of exposure bias is not uniquely supported. Please run a multi-step baseline with the same task-specific loss, or otherwise decouple the loss change from the step-count change.","section":"Table 6, Sec. 3.4"},{"comment":"The semantic enhancer is a load-bearing component (Table 6, last row), but the paper never states where the semantic ground-truth images y_s come from for TransCG, ClearPose, and DREDS/STD, and it reports no semantic segmentation metric (e.g., mIoU) or any predicted semantic map. Given that the training loss in Eq. (8) is an L1 regression on palette-colored images rather than a classification loss, the claim that the model 'jointly performs semantic segmentation' is not verifiable as presented. Please specify the semantic-label source/annotation protocol and report semantic accuracy or at least qualitative semantic outputs on the benchmark splits.","section":"Sec. 3.3, Eq. (8)"}],"minor_comments":[{"comment":"All quantitative claims are reported as single numbers without error bars; for the robotic grasping results, the number of trials and standard deviations should also be reported.","section":"Tables 2-5, Table 8"},{"comment":"The experimental setup for Fig. 4 is underspecified: it should state which model and training loss are used, and whether each curve is an average over multiple noise samples.","section":"Fig. 4"},{"comment":"The notation f_theta([z_x, z_d], m_d) omits the noisy latent z_t and timestep t that appear in main-text Eq. (3); please clarify the relationship between these two notations.","section":"Supplementary Eq. (9)"},{"comment":"The limitations section only mentions efficiency; it should also discuss the sensitivity to the random input noise at t=T and the dependence on semantic ground-truth labels.","section":"Appendix D"},{"comment":"Reference [9] has an empty author field ('Hyojun Go, , Yunsung Lee'), and Table 5 uses inconsistent capitalization ('Didsee' vs. 'DidSee').","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is not circular: the results are measured against external benchmarks and strong baselines, and the core design choices are adaptations of prior work. The main risks are the sign error in the v-prediction reconstruction formula, the potential seed dependence due to the uninformative noise input, and the confounded ablation. These are addressable in revision, so I recommend major revision rather than rejection. I would ask the authors to provide code or a detailed noise-invariance analysis during the revision process."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, practical depth-completion paper for non-Lambertian objects. The central claim—DidSee beats TODE-Trans and others on TransCG, ClearPose, and DREDS/STD—is likely true. The novelty is mostly in the combination rather than any single component, but the combination works and the experiments are thorough.\n\nWhat's new: the rescaled zero-SNR scheduler, the single-step noise-agnostic training at t=T with v-target equal to z0, and the semantic enhancer via task switchers and cross-task attention. Each piece is adapted from prior work ([19], [8], [20], [21]), but putting them together for transparent and specular depth completion is new. The paper does a good job ablating each component (Table 6), showing consistent gains, and the downstream pose and grasp improvements add practical value. The writing is clear and the citation pattern is honest.\n\nSoft spots, in order of softness. First, no error bars anywhere, and no released code. For a method whose inference starts from pure Gaussian noise (alpha_T=0), that is a real omission. The stress-test note is worth taking seriously: the model must learn to ignore the random input latent, and nothing in the paper demonstrates stability across noise samples. One seed could produce a different table. This is the kind of thing that makes a reviewer nervous, even if in practice the strong RGB and raw-depth conditioning probably makes the noise sensitivity small. Second, the single-step ablation is confounded: Table 6 introduces single-step training together with the pixel-space task-specific loss, so the REL drop from 0.020 to 0.017 is not cleanly attributable to the single-step formulation. The authors should have added a variant that isolates the loss change. That is a legitimate weakness, but not a fatal one—the overall SoTA results do not collapse if that component's individual contribution is smaller. Third, the 'systematic analysis' of signal leakage and exposure bias is really just one RMSE-versus-steps curve (Fig. 4). Calling it systematic is a bit much.\n\nBottom line: the paper is worth a serious referee. It should probably be conditional on the authors reporting variance across noise seeds and ideally releasing code, plus an ablation that separates the single-step change from the loss change. I would cite it if I were working on non-Lambertian depth completion; the SoTA numbers are useful even before verification lands.","headline":"A practical, well-validated diffusion-based depth completion system for non-Lambertian objects; the SoTA claims are likely true, but missing error bars and a clean ablation make verification necessary.","tokens_in":16578,"tokens_out":3431,"would_cite":true,"duration_ms":37009,"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":"DidSee turns noisy RGB-D into accurate depth for transparent and specular objects by making the diffusion final step noise-only, predicting depth in one pass, and adding a semantic segmenter, beating prior methods on three benchmarks.","keywords":["depth completion","transparent objects","non-Lambertian surfaces","diffusion models","zero terminal SNR","single-step sampling","semantic segmentation","robotic grasping"],"falsifier":"Run the trained DidSee at timestep $t=T$ with several different Gaussian noise draws and compare the decoded depth maps; the paper's formulation predicts $\\hat{v}_T = \\hat{\\mathbf{z}}^y_0$ independent of $\\boldsymbol{\\epsilon}$, so if the outputs vary noticeably across noise seeds, the noise-agnostic single-step premise is false. A second check is to train with a nonzero terminal $\\bar{\\alpha}_T$ using otherwise identical settings and show that errors do not rise, which would undermine the signal-leakage explanation.","tokens_in":15595,"feed_emoji":"🤖","tokens_out":7360,"duration_ms":74153,"temperature":0.7,"pith_summary":"DidSee sets out to show that a pre-trained text-to-image diffusion model can be turned into an accurate depth-completion engine for transparent and specular objects, the surfaces where RGB-D cameras fail most. The paper argues that the vanilla diffusion setup silently injects two biases — a faint leftover of the clean depth signal at the final noise step, and errors that grow during multi-step sampling — and that both badly hurt depth accuracy even though they matter little for image generation. To remove them, DidSee rescales the noise schedule so the final step is pure noise, then trains the denoiser for a single step at that final timestep, so the network must produce the full depth latent in one pass. A semantic enhancer adds a parallel segmentation task so the model separates object from background where visual features are missing. If the paper's results hold, depth maps restored this way are good enough to improve category-level pose estimation and robotic grasping on materials ranging from diffuse to transparent.","feed_headline":"One-step diffusion restores depth for glass and metal","feed_subtitle":"DidSee removes two diffusion training biases so robots get usable depth on transparent and shiny objects.","key_machinery":"The load-bearing mechanism is the zero-terminal-SNR, noise-agnostic single-step formulation. With the rescaled scheduler, the final forward sample is $\\mathbf{z}^y_T = \\boldsymbol{\\epsilon}$ (since $\\bar{\\alpha}_T = 0$), and the $v$-prediction target in Eq. (7) collapses from $v_T = \\sqrt{\\bar{\\alpha}_T}\\,\\boldsymbol{\\epsilon} + \\sqrt{1-\\bar{\\alpha}_T}\\,\\mathbf{z}^y_0$ to $v_T = \\mathbf{z}^y_0$, so the denoiser's output is directly the clean depth latent and is independent of the sampled noise. This removes exposure bias by training and evaluating at the same single timestep, and it lets the loss operate in pixel space. The semantic enhancer is a second mechanism: discrete segmentation labels are turned into RGB palette images so the VAE can encode them, and two task switchers plus cross-task self-attention let depth and semantic predictions share features, which the paper says gives sharper depth where transparent surfaces blend into backgrounds.","core_discovery":"On the paper's own terms, the central discovery is that the two training-inference mismatches inherited from image-generation diffusion — signal leakage bias and exposure bias — are first-order obstacles to depth completion, not minor artifacts, and that both can be eliminated jointly. DidSee enforces a zero terminal signal-to-noise ratio by rescaling the noise schedule, which makes the last forward step pure Gaussian noise; at that timestep the v-prediction target reduces to the clean depth latent itself, so the model is trained to map raw RGB-D latents plus noise directly to the depth latent in a single forward pass. Because the output is now the predicted clean latent, the model can be supervised with an L1 loss in pixel space instead of latent space. The paper further claims that a semantic enhancer operating through task switchers and cross-task self-attention lets one U-Net predict depth and semantic maps jointly, sharpening boundaries on transparent surfaces. Across TransCG, ClearPose, and DREDS/STD, DidSee reports the lowest errors among the compared methods, and restored depth from DidSee yields higher pose-estimation accuracy and grasping success than depth from earlier completers.","pith_inferences":["The noise-agnostic single-step trick is not obviously limited to depth: any dense regression task that can be encoded in VAE latent space, such as surface normals or specular-reflection masks, could in principle use the same zero-terminal-SNR one-step training to avoid bias accumulation, though the paper does not test this.","If the single-step formulation truly makes output independent of the sampled noise, the model behaves more like a deterministic conditional generator than a diffusion sampler, and one could expect test-time ensembling over noise seeds to yield little or no gain; this is a measurable prediction the paper does not report.","The semantic enhancer's gain on transparent objects suggests that fusing a classification-style task into a regression diffusion model helps exactly where local RGB cues are ambiguous, so task switchers plus cross-task attention could be reused as a general module for other material-agnostic perception tasks."],"forward_implications":["On TransCG, DidSee reports RMSE 0.011 versus 0.013 for TODE-Trans, with relative and mean absolute errors roughly a quarter lower, so even the strongest prior transparent-object completer is beaten by a clear margin.","On ClearPose's hardest scenes, including translucent cover and heavy occlusion, DidSee keeps the lowest errors and its accuracy thresholds stay well above the compared methods, implying the gains are largest exactly where visual cues fail.","Trained only on the synthetic DREDS-CatKnown split, DidSee remains competitive on real STD data and on novel object categories, evidence that the diffusion priors transfer without extra real training data.","Depth from DidSee improves category-level pose estimation over raw or completed depth in most reported metrics, and raises robotic grasping success on transparent, specular, and diffuse objects (overall 92.8% versus 89.2% for TODE-Trans).","Because inference is a single step, the method avoids the multi-step error accumulation shown in the paper and uses one denoising pass plus one VAE decode at test time."],"supporting_citations":[{"why":"Supplies the rescaled noise scheduler and zero-terminal-SNR concept that DidSee adopts to eliminate signal leakage.","marker":"[19]"},{"why":"Identifies the signal-leak bias in diffusion models, the phenomenon DidSee removes for depth completion.","marker":"[4]"},{"why":"Provides the pre-trained Stable Diffusion model and latent VAE that serve as DidSee's backbone.","marker":"[25]"},{"why":"Gives the v-prediction formulation used to derive the single-step target at the terminal timestep.","marker":"[27]"},{"why":"Formalizes exposure bias in diffusion models, the multi-step error accumulation DidSee avoids via single-step training.","marker":"[22]"},{"why":"Contributes the cross-task self-attention mechanism DidSee uses to let depth and semantic features guide each other.","marker":"[8]"},{"why":"Contributes the task-switcher conditioning used to make one U-Net perform depth and semantic regression jointly.","marker":"[20]"},{"why":"Provides the color-palette encoding that turns discrete semantic labels into RGB images processable by the VAE.","marker":"[32]"},{"why":"Supplies the TransCG dataset, the primary real-world training and evaluation benchmark for transparent-object depth completion.","marker":"[5]"},{"why":"Supplies the DREDS and STD datasets and the SWinDRNet evaluation protocol used for sim-to-real and pose-estimation tests.","marker":"[3]"}],"fun_headline_variants":["Diffusion depth fix for shiny and clear objects","DidSee: one-step depth from noisy RGB-D","Zero SNR schedule sharpens robot depth sensing","Diffusion bias removal yields precise depth maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one forward pass of the U-Net, starting from pure Gaussian noise plus the RGB and raw-depth latents, can directly produce a depth latent that the frozen VAE decodes into accurate depth; if the network cannot learn that underdetermined mapping, or if the random input noise makes predictions unstable, the central performance claims collapse.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion depth fix for shiny and clear objects","DidSee: one-step depth from noisy RGB-D","Zero SNR schedule sharpens robot depth sensing","Diffusion bias removal yields precise depth maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1263,"prompt_tokens":999,"completion_tokens":264,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":218}},"tokens_in":615,"tokens_out":264,"duration_ms":3632,"temperature":1.0,"reasoning_tokens":218,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:35:06.641957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained DidSee at timestep $t=T$ with several different Gaussian noise draws and compare the decoded depth maps; the paper's formulation predicts $\\hat{v}_T = \\hat{\\mathbf{z}}^y_0$ independent of $\\boldsymbol{\\epsilon}$, so if the outputs vary noticeably across noise seeds, the noise-agnostic single-step premise is false. A second check is to train with a nonzero terminal $\\bar{\\alpha}_T$ using otherwise identical settings and show that errors do not rise, which would undermine the signal-leakage explanation.","supporting_citations":[{"cited_title":"Common diffusion noise schedules and sample steps are flawed","cited_arxiv_id":null,"evidence_quote":"Supplies the rescaled noise scheduler and zero-terminal-SNR concept that DidSee adopts to eliminate signal leakage."},{"cited_title":"Exploiting the signal-leak bias in diffusion models","cited_arxiv_id":null,"evidence_quote":"Identifies the signal-leak bias in diffusion models, the phenomenon DidSee removes for depth completion."},{"cited_title":"High-resolution image syn- thesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the pre-trained Stable Diffusion model and latent VAE that serve as DidSee's backbone."},{"cited_title":"Progressive distillation for fast sampling of diffusion models","cited_arxiv_id":null,"evidence_quote":"Gives the v-prediction formulation used to derive the single-step target at the terminal timestep."},{"cited_title":"Elucidating the exposure bias in diffusion models","cited_arxiv_id":null,"evidence_quote":"Formalizes exposure bias in diffusion models, the multi-step error accumulation DidSee avoids via single-step training."},{"cited_title":"Geowiz- ard: Unleashing the diffusion priors for 3d geometry estima- tion from a single image","cited_arxiv_id":null,"evidence_quote":"Contributes the cross-task self-attention mechanism DidSee uses to let depth and semantic features guide each other."},{"cited_title":"Wonder3d: Single image to 3d using cross-domain diffusion","cited_arxiv_id":null,"evidence_quote":"Contributes the task-switcher conditioning used to make one U-Net perform depth and semantic regression jointly."},{"cited_title":"Unsupervised modality adapta- tion with text-to-image diffusion models for semantic seg- mentation","cited_arxiv_id":null,"evidence_quote":"Provides the color-palette encoding that turns discrete semantic labels into RGB images processable by the VAE."},{"cited_title":"Transcg: A large-scale real-world dataset for transparent object depth completion and a grasping baseline","cited_arxiv_id":null,"evidence_quote":"Supplies the TransCG dataset, the primary real-world training and evaluation benchmark for transparent-object depth completion."},{"cited_title":"Domain randomization- enhanced depth simulation and restoration for perceiving and grasping specular and transparent objects","cited_arxiv_id":null,"evidence_quote":"Supplies the DREDS and STD datasets and the SWinDRNet evaluation protocol used for sim-to-real and pose-estimation tests."}],"review_version":1}