{"id":"f575b838-3647-41f9-a309-6ece43e5f897","arxiv_id":"2411.18229","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A ground-truth-free diffusion distillation method makes metric depth predictions from UniDepth sharper while preserving their scale.","lead":"SharpDepth refines monocular metric depth estimates from models like UniDepth by using a diffusion model to sharpen object boundaries while keeping the absolute scale. It trains without ground-truth depth on about 90,000 real images, using two pretrained depth models as teachers, and reports improved boundary detail on zero-shot benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper's own DBE/PDBE tables contradict the sharpness claim: SharpDepth has worse boundary completion than UniDepth-aligned Lotus on all four detail benchmarks.","rationale":"The paper's contribution is the claim of combining metric accuracy with generative sharpness. The reader's weakest assumption focused on metric accuracy being inherited from UniDepth; I agree with that concern, but the more decisive problem is in the sharpness evidence itself. The baseline UniDepth-aligned Lotus is exactly the control that isolates the learned refinement: take the generative depth prior and align it to the metric model. If SharpDepth cannot beat that control on the sharpness metrics, then the only measured benefit of the learned network over trivial alignment is in A.Rel, which is confounded by the test-time least-squares alignment to UniDepth. Table 2 shows that SharpDepth's DBE/PDBE completion is worse than this control on all four datasets; accuracy is mixed. Thus the reported numbers support a trade-off, not the simultaneous gain stated in the abstract. This is an internal inconsistency, not a dispute with external consensus. The proposed check (compute official iBims DBE and an aggregate boundary score) would settle whether the paper's chosen metrics were misleading; if the ordering holds, the central claim is unsupported and the paper should be rejected or substantially revised. My verdict therefore moves from the reader's CONDITIONAL to REJECT.","tokens_in":16260,"tokens_out":10640,"duration_ms":96725,"concrete_test":"Obtain the authors' predicted depth maps (or rerun their released code) for SharpDepth and UniDepth-aligned Lotus on iBims, and compute the official iBims DBE with the standard implementation, reporting both accuracy and completion plus a combined boundary F-score. If SharpDepth still has worse completion or worse aggregate boundary quality than UniDepth-aligned Lotus, the sharpness claim fails. As a complementary check, recompute PDBE on the three synthetic sets with the same Canny thresholds used by the authors to verify the reported ordering; if the ordering flips under reasonable threshold choices, the reported sharpness advantage is a threshold artifact.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that SharpDepth yields metric depth that is both accurate and sharp, recovering fine structures while keeping UniDepth's scale. The paper's own detail benchmark (Tab. 2) does not support the 'sharp' half. On the completion component of DBE/PDBE (lower is better), SharpDepth is worse than the UniDepth-aligned Lotus baseline on every reported dataset: Sintel 36.2 vs 31.9, UnrealStereo4K 61.5 vs 33.2, Spring 147.6 vs 102.7, iBims 13.1 vs 11.0. It is also worse than GT-aligned Lotus on all four. UniDepth-aligned Lotus is the natural control: it is the generative prior (Lotus) rescaled to UniDepth without any learned refinement. On the accuracy component, SharpDepth is better on three of four datasets (Sintel 1.94 vs 2.04, Spring 1.24 vs 1.27, iBims 1.80 vs 1.92) and worse on UnrealStereo4K (1.37 vs 1.21), so the evidence is mixed. The only consistent quantitative advantage of SharpDepth over this baseline is metric accuracy (A.Rel), which is expected from the test-time least-squares alignment to UniDepth described in Sec. 5.1. The paper's own numbers therefore show a trade-off, not the claimed simultaneous gain: alignment to UniDepth buys metric accuracy, and the learned refinement loses boundary completeness relative to trivial alignment. This is load-bearing: without a demonstrated sharpness gain over the aligned baseline, the method is not an advance over a one-line post-processing baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"SharpDepth proposes a diffusion-based depth sharpener that refines the metric depth predictions of a discriminative model (UniDepth) using the generative prior of an affine-invariant diffusion depth model (Lotus). The method is trained without ground-truth depth, using a difference map between the metric and affine-invariant predictions to drive a noise-aware gating mechanism, an SDS distillation loss, and a noise-aware reconstruction loss. The paper reports zero-shot depth accuracy on seven real datasets and boundary detail metrics on three synthetic and one real dataset, claiming that SharpDepth achieves both metric accuracy and sharp boundaries. The central evidence is Tables 1 and 2, with UniDepth-aligned Lotus introduced as a baseline that rescales Lotus to UniDepth using scale and shift.","tokens_in":16584,"tokens_out":4426,"duration_ms":41913,"significance":"If the claims were fully supported, the paper would make a useful contribution: a ground-truth-free, data-efficient refinement of metric depth that adds high-frequency detail while preserving absolute scale has clear practical value, and the proposed noise-aware gating and EMA-based teacher updating are reasonable design ideas. The training scheme on roughly 90k images is substantially cheaper than typical discriminative depth pipelines, which is a concrete strength. However, the evaluation as presented does not establish the central claim of simultaneous accuracy and sharpness: the sharpness metrics in Table 2 do not show an advantage over the simple UniDepth-aligned Lotus baseline, and the metric accuracy comparison is confounded by test-time least-squares alignment to UniDepth. The paper's contribution is therefore not yet convincingly demonstrated.","major_comments":[{"comment":"The final output is post-processed with least-squares alignment to the original UniDepth depth, restricted to pixels with minimal differences. This means the reported A.Rel, RMSE, and delta1 values in Tables 1, 2, and 6 are not the raw output of the learned model; part of the metric accuracy is enforced by construction. The claim that SharpDepth 'keeps the scale of UniDepth' is trivial under this protocol. The paper should report metrics without this test-time alignment, or alternatively show that the raw output is already metrically accurate and justify the alignment as an independent post-processing step. Without this, the consistent A.Rel advantage over UniDepth-aligned Lotus may be an artifact of the more refined alignment procedure rather than of the learned refinement.","section":"Sec. 5.1, Implementation details"},{"comment":"The paper's own detail benchmark contradicts the sharpness half of the central claim. On the completion component of PDBE/DBE, SharpDepth is worse than UniDepth-aligned Lotus on all four reported datasets: Sintel 36.2 vs. 31.9, UnrealStereo4K 61.5 vs. 33.2, Spring 147.6 vs. 102.7, and iBims 13.1 vs. 11.0. Since UniDepth-aligned Lotus is the natural control (the generative prior rescaled to UniDepth without any learned refinement), the learned pipeline does not demonstrate improved boundary completeness; it is worse. The sentence in Sec. 5.2 that SharpDepth 'consistently surpass[es] the naive UniDepth-aligned Lotus baseline across all datasets' is only true for depth accuracy metrics, not for boundary completion. The authors should either provide a boundary metric on which SharpDepth clearly beats this baseline, or revise the claim to describe a trade-off rather than a simultaneous gain.","section":"Table 2, PDBE/DBE completion"},{"comment":"The gating mechanism and the test-time alignment both rely on the assumption that pixels where UniDepth and Lotus agree are reliable and need no refinement. This assumption can fail when the two models share systematic errors, for example on thin structures, transparent surfaces, or specular regions, and the paper provides no validation that low-difference pixels actually agree with ground truth. This is load-bearing for the method, because both the noise-aware gating and the reconstruction loss direct the model away from exactly those regions. The authors should add an analysis (e.g., on a small labeled validation set) showing that low-difference pixels are indeed trustworthy, or discuss the failure mode and why it is not fatal.","section":"Sec. 4.1, difference map assumption"}],"minor_comments":[{"comment":"The PDBE metric is introduced but its accuracy and completion components are not defined in the main text; a brief definition or a pointer to the supplementary equation would help readers interpret Table 2.","section":"Sec. 5.1, Metrics"},{"comment":"The Pareto-frontier plot aggregates DBE Completion over three synthetic datasets, but the aggregation rule is not stated; please specify whether it is a mean over Sintel, UnrealStereo4K, and Spring, and whether iBims is included.","section":"Figure 2"},{"comment":"The SDS loss in Eq. (5) omits the weighting w_t that appears in Eq. (3); please state the exact weighting and noise schedule used in the implementation.","section":"Eq. (5)"},{"comment":"The norm in Eq. (6) is unspecified; if it is the L1 norm, this should be stated, as the type of norm affects the optimization behavior.","section":"Eq. (6)"},{"comment":"The phrase 'the output is rescaled using least-squares alignment with the original metric depth' is ambiguous: it should be stated whether the least-squares fit uses only low-difference pixels or weights them, and how the difference map threshold is determined.","section":"Sec. 5.1, Implementation details"},{"comment":"The claim that 90,000 images is '100-150 times smaller than the amount of data used for discriminative depth models' needs an explicit comparison; the reference models and their training set sizes should be named.","section":"Sec. 5.1, Training data size"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the training scheme is data-efficient, but the evaluation currently does not support the advertised claim of simultaneous metric accuracy and boundary sharpness. The most serious issue is the absence of any sharpness gain over the UniDepth-aligned Lotus baseline on the paper's own boundary-completion metric, combined with the test-time least-squares alignment that partially manufactures the metric accuracy. I would recommend asking the authors to report unaligned metrics, to benchmark against UniDepth-aligned Lotus on boundary metrics in a controlled way, and to either substantiate or soften the central claim. If the sharpness gain cannot be shown, the paper should be repositioned as a metric-alignment post-processor or an accuracy-refinement method, rather than a sharpening method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the training recipe is new and worth engaging, but the headline claim is not supported by the paper's own tables. The stress-test note holds up on inspection: on the completion component of their boundary metric, SharpDepth is worse than the UniDepth-aligned Lotus baseline on all four datasets (Sintel 36.2 vs 31.9, UnrealStereo4K 61.5 vs 33.2, Spring 147.6 vs 102.7, iBims 13.1 vs 11.0). That baseline is the right control—the generative prior rescaled to UniDepth without learned refinement—so the 'sharp' half of the abstract is directly contradicted. The metric-accuracy advantage over that baseline is also mostly a byproduct of the test-time least-squares alignment to UniDepth on low-difference pixels, which the paper states in Sec. 5.1. So the method delivers a trade-off, not the simultaneous gain advertised.\n\nWhat is genuinely new: the combination of SDS distillation, a noise-aware gating mechanism, and a difference-map-weighted reconstruction loss to train a refiner on real images with no ground truth. The ablations are informative—settings A–H confirm that each component contributes and that the EMA teacher helps. The writing is clear, and I credit the authors for including the aligned-Lotus baseline; they just do not let its numbers temper their claims.\n\nSoft spots, in proportion: (1) The evaluation is partially self-referential. Aligning the output to UniDepth at test time and then reporting accuracy versus UniDepth makes the comparison circular; reporting results without that final alignment, or holding alignment fixed across methods, would be more convincing. (2) The PDBE metric is self-defined, and the paper reports no error bars and releases no code, so the numbers cannot be checked or variance assessed. (3) The sharpness claim rides on the edge-accuracy component (where SharpDepth wins three of four) while completion is worse everywhere; the text should acknowledge this trade-off explicitly. That is load-bearing for the paper's stated contribution.\n\nBottom line: this deserves a serious referee. The training recipe is interesting and the zero-shot evaluation is broad, but the paper needs a major revision to decouple the alignment effect from the learned refinement, report variance, and recalibrate the claims. I would send it to review with a request to fix the evaluation rather than desk-reject it.","headline":"SharpDepth has a plausible training recipe, but its own boundary-completion numbers show it losing to the trivial UniDepth-aligned Lotus baseline, so the simultaneous accuracy-and-sharpness claim doesn't hold as stated.","tokens_in":17113,"tokens_out":3995,"would_cite":false,"duration_ms":34054,"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":"SharpDepth claims a diffusion-based refiner can add sharp boundaries to metric depth predictions while preserving absolute scale, using no ground-truth depth during training.","keywords":["monocular metric depth estimation","diffusion models","depth refinement","zero-shot depth estimation","score distillation sampling","depth boundary sharpness","difference map gating","ground-truth-free training"],"falsifier":"Run SharpDepth on a domain where UniDepth is known to be systematically wrong—for example, highly reflective or transparent surfaces such as those in the Booster benchmark—and compare its aligned RMSE against UniDepth's; if SharpDepth's error tracks UniDepth's error one-to-one and no shared error is ever corrected, the claim that the refiner adds independent metric accuracy is refuted.","tokens_in":16073,"feed_emoji":"📏","tokens_out":8080,"duration_ms":64506,"temperature":0.7,"pith_summary":"SharpDepth claims that the two families of monocular depth estimators—discriminative models that predict accurate metric depth but smooth over fine structure, and generative diffusion models that predict sharp relative depth but lack scale—can be combined into a single zero-shot metric depth network. The paper builds a diffusion-based refiner on top of an existing metric depth model and an affine-invariant diffusion depth model, and trains it with no ground-truth depth at all. The reported results show accuracy comparable to the metric backbone while depth-boundary errors drop substantially, so thin structures like fences and poles survive in the final depth map. If the claim holds, high-quality metric depth for AR, robotics, and reconstruction can be obtained from a small, label-free training set instead of millions of annotated or synthetic images.","feed_headline":"Diffusion refiner adds sharp edges to metric depth maps","feed_subtitle":"Metric scale stays intact while diffusion-style sharpness is added, with no ground-truth labels.","key_machinery":"The load-bearing mechanism is Noise-aware Gating combined with the difference map. The difference map $e$ is computed between the normalized UniDepth and Lotus predictions; the latent of the metric depth is then blended as $z'_d = \\hat{e} \\odot \\epsilon + (1-\\hat{e}) \\odot z_d$, so high-disagreement regions receive more noise and become the focus of the diffusion sharpener. Two losses drive training: a modified Score Distillation Sampling (SDS) loss distills fine detail from the pretrained diffusion depth model, and a Noise-aware Reconstruction loss $L_{\\text{recons}} = \\| e \\odot (\\hat{d} - d) \\|$ keeps the output anchored to the metric estimate. An exponential moving average of the training model replaces the static teacher as training progresses, so the difference map shrinks as the sharpener improves.","core_discovery":"The central claim is that a diffusion-based depth sharpener can refine the output of a metric depth estimator so that the final prediction is simultaneously metrically accurate and visually sharp, without any ground-truth depth supervision. SharpDepth takes the metric depth map from UniDepth and the affine-invariant depth map from Lotus, normalizes both to a common range, and forms a difference map that marks where the two models disagree. The refiner receives a latent in which disagreement regions are heavily noised and agreement regions are mostly preserved, then is trained with a score distillation loss from the diffusion depth teacher and a noise-aware reconstruction loss that keeps the output close to the metric estimate. At test time the decoded output is rescaled by least-squares alignment to the original metric depth on low-difference pixels. The paper reports that this restores sharp boundaries and thin structures while retaining the absolute scale of the metric backbone.","pith_inferences":["The method's metric accuracy is effectively inherited: the final least-squares rescale to UniDepth means SharpDepth can be no more metric than UniDepth on the aligned pixels, and a better or worse metric teacher would shift the reported accuracy accordingly.","The difference-map trust rule assumes that agreement between UniDepth and Lotus marks trustworthy pixels, so errors the two models share—for example, systematic biases from similar training data—would be invisible to the gating and passed through.","A natural extension is to treat the metric backbone as a swappable module: the same trained sharpener could be applied to any metric depth model, and the paper's qualitative Metric3Dv2 results suggest this transfer works without retraining.","Because the SDS loss encourages consistency with the diffusion teacher's distribution, the sharpness gain may be limited on scene contents the teacher was never trained on; feeding synthetic or out-of-domain images could test how much of the detail comes from the teacher prior versus the input image."],"forward_implications":["If SharpDepth is correct, metric depth estimators can be sharpened post hoc without re-training on ground-truth depth, making detail recovery cheap.","The training set of roughly 90,000 real images suggests that label-free refinement can reach competitive zero-shot accuracy with 100–150 times less data than discriminative metric models.","Downstream tasks that consume depth maps—point-cloud fusion, TSDF reconstruction, and Gaussian-splatting SLAM—should inherit sharper object boundaries from the improved depth maps, as demonstrated on TUM fr1/desk and KITTI reconstructions.","Because alignment is performed only on low-difference pixels, the final metric scale is tied to the metric backbone, so SharpDepth's accuracy cannot exceed the reliability of that backbone on the aligned pixels."],"supporting_citations":[{"why":"Supplies the metric depth backbone whose scale SharpDepth preserves and to which the final output is aligned.","marker":"[31]"},{"why":"Supplies the affine-invariant diffusion depth prior as well as the architecture and pretrained weights used to initialize the sharpener.","marker":"[17]"},{"why":"Represents the diffusion depth estimation line that Lotus builds on and serves as an alternative teacher in ablations.","marker":"[21]"},{"why":"Source of the Score Distillation Sampling formulation that SharpDepth adapts for distilling depth detail.","marker":"[32]"},{"why":"Informs the variational score distillation approach used to train the one-step sharpener.","marker":"[30]"},{"why":"A metric refinement baseline that SharpDepth compares against and consistently outperforms in accuracy and detail.","marker":"[25]"},{"why":"An existing depth-refinement baseline whose reliance on synthetic data SharpDepth contrasts with its ground-truth-free real-data protocol.","marker":"[55]"},{"why":"A metric depth baseline used for comparison and as the target of the generalization experiment where SharpDepth is applied without retraining.","marker":"[19]"}],"fun_headline_variants":["Diffusion distillation sharpens metric depth maps","Razor-sharp metric depth from diffusion distillation","Metric depth gets sharp edges via diffusion distillation","Diffusion refiner keeps metric depth sharp and scaled","Diffusion distillation sharpens metric depth without ground truth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pixels where UniDepth and Lotus agree are reliable enough to anchor both the difference map and the final least-squares rescale, so any bias shared by the two models is inherited by SharpDepth.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion distillation sharpens metric depth maps","Razor-sharp metric depth from diffusion distillation","Metric depth gets sharp edges via diffusion distillation","Diffusion refiner keeps metric depth sharp and scaled","Diffusion distillation sharpens metric depth without ground truth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00091,"raw_usage":{"total_tokens":3879,"prompt_tokens":885,"completion_tokens":2994,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":2923}},"tokens_in":501,"tokens_out":2994,"duration_ms":20969,"temperature":1.0,"reasoning_tokens":2923,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:23:03.499942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SharpDepth on a domain where UniDepth is known to be systematically wrong—for example, highly reflective or transparent surfaces such as those in the Booster benchmark—and compare its aligned RMSE against UniDepth's; if SharpDepth's error tracks UniDepth's error one-to-one and no shared error is ever corrected, the claim that the refiner adds independent metric accuracy is refuted.","supporting_citations":[{"cited_title":"Repurpos- ing diffusion-based image generators for monocular depth estimation","cited_arxiv_id":null,"evidence_quote":"Represents the diffusion depth estimation line that Lotus builds on and serves as an alternative teacher in ablations."},{"cited_title":"Swiftbrush: One-step text-to-image diffusion model with variational score distilla- tion","cited_arxiv_id":null,"evidence_quote":"Informs the variational score distillation approach used to train the one-step sharpener."},{"cited_title":"Patchre- finer: Leveraging synthetic data for real-domain high- resolution monocular metric depth estimation","cited_arxiv_id":null,"evidence_quote":"A metric refinement baseline that SharpDepth compares against and consistently outperforms in accuracy and detail."}],"review_version":1}