{"id":"494b54ec-088e-4fed-9f87-b1334df48837","arxiv_id":"2508.13091","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DMS fine-tunes Stable Diffusion to synthesize left-left, right-right, and center views from stereo images, then uses these synthetic views in a per-pixel minimum warping loss to reduce outliers in self-supervised depth estimation.","lead":"The paper proposes DMS, a pipeline that fine-tunes Stable Diffusion to generate extra synthetic views around a stereo pair, then uses those views as additional photometric supervision for self-supervised depth estimation. The method reports consistent benchmark improvements, but the headline 'cost-free' claim hides substantial compute and the geometric reliability of the generated views is only loosely verified.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed geometric consistency of the generated center view rests on an unverified scale-invariance assumption; PSNR/SSIM on CARLA does not establish that its disparity is half the true left-right disparity.","rationale":"The reader's weakest assumption identifies the same load-bearing step: the center view's geometric consistency depends on an empirical, under-validated scale-invariance property of the fine-tuned diffusion model. This is not just a missing ablation; it is the mechanism by which the method claims to provide explicit photometric correspondences for occluded and out-of-frame regions. The paper's existing CARLA evaluation measures PSNR/SSIM, which can be high for plausible textures even when the underlying disparities are wrong. The per-pixel minimum loss makes the issue more subtle, since min over multiple warped views can mask errors in individual views. I nevertheless do not think this warrants REJECT: the method is clearly specified, the code is promised, the ablation tables are internally consistent, and the concern is addressable by a direct disparity measurement on CARLA. The appropriate verdict remains CONDITIONAL; no adjustment to the reader's verdict is needed.","tokens_in":22088,"tokens_out":7447,"duration_ms":82512,"concrete_test":"On the CARLA multi-baseline dataset (which has GT for left, center, right, left-left, right-right), run the DMS generator with X=2.0 to create a center view from the left image, then estimate per-pixel disparity between the generated center view and the true left view using the GT depth (or a strong stereo/optical-flow model), and compare it with 0.5*d_GT. Report mean absolute error and the fraction of pixels with error >1px. If the generated center disparity deviates systematically from 0.5*d_GT (or if the same check fails for the ll/rr views with their expected scale factors), then the scale factors in Eq. 5 are not justified and the claimed geometric mechanism is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive step is the intermediate view approximation in Sec. 3.2.3 / Fig. 3: upscaling the input by X=2.0 and applying the learned left-to-right generator is claimed to yield a valid center view with disparity scale s_c=0.5 in Eq. 5. This rests entirely on the empirical observation that 'upscaling the input resolution yields the same pixel-level displacements in generated views.' For a true center camera at half baseline, per-pixel disparity must equal d_GT/2. The proposed operation yields that only if the diffusion model's output displacement is invariant to input resolution and equals the original disparity field; neither is derived, and the only quantitative evidence is PSNR/SSIM on CARLA (Tables 2, 10), which does not measure geometric correctness. The same gap applies to the left-left and right-right views, which extrapolate beyond the left-to-right training distribution and are validated only by appearance metrics. If the generated views have systematic disparity errors, the per-pixel minimum warping loss in Eq. 5 can still lower photometric error by selecting the view whose artifacts best match the target, so the EPE/D1 gains in Tables 5-6 would not establish correct geometric supervision. A direct disparity check is therefore the load-bearing missing experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DMS, a two-stage method for self-supervised depth estimation. In stage 1, a Stable Diffusion V2 model is fine-tuned on unlabeled stereo pairs with direction prompts ('to left'/'to right') to synthesize multi-baseline views: left-left, right-right, and a center view obtained by upscaling the input by a factor X before generation. In stage 2, these views are incorporated into the photometric training of self-supervised stereo matching and stereo-supervised monocular depth networks through a per-pixel minimum warping loss (Eq. 5) with fixed disparity scale factors. Experiments on SceneFlow, KITTI, and MPI-Sintel show consistent EPE/D1 improvements, including up to about 35% outlier reduction on the KITTI 2015 test set, and the method is demonstrated across several stereo and monocular backbones.","tokens_in":22292,"tokens_out":6347,"duration_ms":66696,"significance":"If the geometric consistency of the generated views were established, DMS would be a practically valuable and unusually simple contribution: it turns a large pretrained diffusion model into a multi-baseline stereo generator without extra annotations, and it improves several self-supervised depth baselines in a model-agnostic way. The paper's strengths include the large-scale evaluation across datasets and network architectures, the inclusion of ablations that separate the contributions of left-left/right-right and center views, and the public code release. The main open risk is that the central geometric claim is asserted rather than verified; the currently reported appearance metrics are not sufficient to support it.","major_comments":[{"comment":"The generation of the center view rests on an unverified scale-invariance assumption. The text states that \"upscaling the input resolution yields the same pixel-level displacements in generated views,\" but the paper provides neither a derivation nor a direct measurement of the generated disparity. The CARLA evaluation in Table 10 reports only PSNR and SSIM, which do not establish that the center view's disparity is d/2 or that the left-left and right-right views have disparities of -d and 2d. Since Eq. (5) uses the fixed scales s_c=0.5, s_ll=-1, and s_rr=2, an incorrect scale factor would allow the min operator to select the view whose artifacts best match the target and still reduce the photometric loss, so the EPE/D1 gains in Tables 5 and 6 would not demonstrate correct geometric supervision. Please add a direct disparity-level evaluation on CARLA (e.g., EPE/D1 of the generated views against the simulator ground truth, or per-view warping error under GT disparity) and report the actual measured displacement scale for each generated view.","section":"Section 3.2.3 and Eq. (5)"},{"comment":"The benchmark comparison may not isolate the effect of DMS. For the learning-based entries marked with an asterisk, both the baseline and +DMS numbers are re-implemented \"after unsupervised training with vanilla warping loss,\" but the paper does not state whether the two rows use identical training budgets, augmentations, hyperparameters, and random seeds, nor does it report variance or standard error for the KITTI test results. Since several methods show more than 30% D1-bg reduction, a small difference in the fine-tuning protocol could explain part of the gain. Please specify the exact protocol and, ideally, report the re-implemented baseline alongside the original published numbers so the reader can verify that the comparison is controlled.","section":"Section 4.3.2 and Table 6"},{"comment":"The occluded-region evaluation on KITTI relies on pseudo-occlusion masks produced by a pre-trained stereo model rather than ground-truth occlusions, while the text says \"ground disparity\" is used for the mask formula. This makes the reported \"improvement in occluded regions\" on KITTI harder to interpret, as errors in the pseudo masks are correlated with the quality of the stereo method being evaluated. Please clarify the exact mask source for each dataset in Table 5 and, if possible, report the occluded-region numbers on SceneFlow and MPI-Sintel, where ground-truth occlusion masks are available, as the primary evidence.","section":"Supplementary Section 5.1.3 and Table 5"}],"minor_comments":[{"comment":"The weighting parameter alpha in the photometric loss is never defined in the main text; please state its value or reference where it is set.","section":"Section 3.3, Eq. (6)"},{"comment":"The sentence referring to \"Equation 3.3\" is ambiguous; it should reference the per-pixel minimum warping loss (Eq. 5) or the exact equation used for the SGM cost volume construction.","section":"Section 4.3.2"},{"comment":"The abstract and introduction call DMS \"cost-free,\" but Table 8 reports 4-6 seconds per image and about 7 GB GPU memory for inference; please qualify the term, for example by stating that no additional labels or network modifications are required.","section":"Section 4.1 and Table 8"},{"comment":"There is a typo in the mask generation description: \"pseudo-left and pseudo-left disparities\" should presumably read \"pseudo-left and pseudo-right disparities\"; please also clarify whether the KITTI occlusion masks used in Table 5 come from this pre-trained model and how its accuracy affects the reported occluded-region gains.","section":"Supplementary Section 5.1.3"},{"comment":"Several typos should be corrected: \"ScenceFlow\" in Section 4.3.1 and Table 5, \"Gemetric Consistency\" in Table 3, and \"Denosing UNet\" in Section 4.1.","section":"Various"},{"comment":"The phrase \"check-marked annotations\" is unclear; state explicitly which rows are novel views generated by the proposed inference procedure rather than the existing left/right views.","section":"Table 10 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is on the right track and the empirical gains are suggestive, but the geometric-consistency claim is the hinge of the method. If the authors add a direct disparity-level validation on CARLA and clarify the controlled comparison in Table 6, the contribution could become publishable; without those additions, the central mechanism remains unverified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is using a fine-tuned Stable Diffusion model with directional prompts to synthesize left-left, right-right, and center views along the epipolar line, then feeding those into a per-pixel minimum warping loss. The experiments are broad and the gains are consistent across SceneFlow, KITTI, MPI-Sintel, and on several stereo and monocular baselines. The plug-and-play integration with SGM, PASMnet, RAFTStereo, IGEVStereo, and MCStereo is well executed, and the ablations clearly show the extra views help, especially in occluded and out-of-frame regions. That is real value.\n\nThe main soft spot is the center view. The method upscales the input by X, runs the left-to-right generator, and assumes the same pixel displacement in the upscaled image, so a 2x upscale yields half disparity. That is a plausible and clever empirical trick, but it is not derived and the paper's only quantitative support is PSNR/SSIM on CARLA. Those appearance metrics do not test whether the generated center view actually has disparity d/2. The warping visualization in Figure 4 is suggestive, but not a measurement. The missing experiment is a direct comparison of the generated center view's disparity against CARLA's ground-truth center image. Without that, some of the +c gain could come from the min-over-more-views behavior selecting the least-bad artifact rather than from geometrically correct correspondences. I do not think this kills the paper, because the ablation pattern and the visualizations are consistent with correct geometry, and the min loss is conservative, but it is a load-bearing assumption that needs to be tested.\n\nThe 'cost-free' label is an overstatement at 4–6 seconds per image on an A6000. A non-diffusion baseline for the intermediate view, such as simple horizontal interpolation, and multiple seeds or variance reporting would also strengthen the evidence. The circularity of fine-tuning the diffusion model on the same stereo pairs used for depth training is real but mild, since the depth network still uses the original photometric loss and is evaluated on real ground truth.\n\nWho is this for: researchers in self-supervised stereo and monocular depth, and anyone using diffusion priors for geometric tasks. It deserves a serious referee; the core idea is novel and the evaluation is strong enough to warrant deeper scrutiny rather than desk rejection.","headline":"A novel diffusion-based multi-baseline view generator that improves self-supervised depth, but the center-view geometry rests on an unverified upscaling assumption and needs a direct disparity check.","tokens_in":22867,"tokens_out":3898,"would_cite":true,"duration_ms":43019,"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":"A diffusion model that synthesizes extra stereo views can cut disparity outliers by up to 35 percent in self-supervised depth estimation.","keywords":["self-supervised depth estimation","stereo matching","diffusion models","novel view synthesis","multi-baseline stereo","photometric warping loss","Stable Diffusion","monocular depth estimation"],"falsifier":"In a CARLA scene with ground-truth multi-baseline views, generate the center view from the left image at X=2 and compute the end-point error between this generated view warped by true disparity and the true center image; if the center view does not sit at half the left-right disparity, the geometric-consistency assumption behind the minimum warping loss fails.","tokens_in":21824,"feed_emoji":"📐","tokens_out":7934,"duration_ms":71863,"temperature":0.7,"pith_summary":"Self-supervised depth estimators trained on stereo pairs struggle where pixels are visible in only one camera — occlusions and out-of-frame areas — because the photometric reconstruction loss has no matching pixel to use. The paper proposes DMS, a plug-in that fine-tunes a latent diffusion model to generate additional views along the epipolar line: a left-left view, a right-right view, and a center view between the two cameras. These synthesized views supply explicit correspondences, so a per-pixel minimum warping loss can pick the source view that actually sees each pixel. Training and synthesis need only unlabeled stereo pairs, and the authors report up to 35% outlier reduction and improved disparity maps across several benchmarks.","feed_headline":"Diffusion-made stereo views cut depth outliers up to 35%","feed_subtitle":"Left-left, right-right, and center views give occluded pixels explicit matches for self-supervised depth.","key_machinery":"The mechanism is the repurposed latent diffusion model used as an epipolar novel-view synthesizer: it encodes the reference image into a latent, concatenates the target latent during training, and is steered at inference by text direction prompts ('to left' / 'to right') to produce left-left and right-right views. An intermediate view is obtained by a resolution-scale trick — upscaling the input by X=2 before generation produces the same pixel displacement in the upscaled coordinates, hence half the disparity at the original resolution. The synthesized views enter the Per-Pixel Minimum Warping Loss, which takes the minimum photometric error over source views {right, left-left, right-right, center} with disparity scale factors {1, -1, 2, 0.5}, so each pixel is matched from whichever generated view actually sees it.","core_discovery":"On its own terms, the paper establishes that a latent diffusion model fine-tuned on ordinary unlabeled stereo pairs can act as a multi-baseline stereo generator: conditioned on one image and a direction prompt ('to left' or 'to right'), it produces geometrically consistent novel views beyond the original pair, and upscaling the input by a factor X=2 yields an intermediate center view with proportionally smaller horizontal displacement. When these views are fed into the per-pixel minimum warping loss alongside the original right image, occluded and out-of-frame regions gain explicit photometric correspondences, improving disparity estimation for self-supervised stereo matching and stereo-supervised monocular depth estimation. The method is model-agnostic and adds no labeled supervision.","pith_inferences":["If the resolution-scaling observation generalizes to arbitrary factors, the same diffusion generator could produce a continuum of baselines rather than fixed units, enabling adaptive baseline selection for different depth ranges; the paper only tests X=2 as the center view.","The same multi-view minimum-warping supervision should transfer to self-supervised optical flow or scene flow, where occlusion creates the identical missing-correspondence problem; the paper does not test those tasks.","A stricter check of geometric consistency would measure the generated center view's disparity against ground truth (e.g., on a multi-baseline simulator) instead of PSNR/SSIM; if the X=2 view deviates from half disparity, photometric matching would bias depth estimates at object boundaries."],"forward_implications":["Adding the generated left-left, right-right, and center views reduces outlier ratios substantially in occluded and out-of-frame regions, for example from 83.6% to 49.2% occlusion outliers on SceneFlow with the baseline stereo network.","DMS improves both classical stereo (SGM) and learning-based stereo networks on the KITTI 2015 benchmark without changing their architectures.","The same generated views improve stereo-supervised monocular depth estimators on the KITTI Eigen split.","Combining all generated views outperforms any single one; the center view alone gives the best intermediate result, and adding it to left-left and right-right yields the largest gains.","Because the method uses only unlabeled stereo pairs for both synthesis and training, it can be applied to any new stereo dataset without extra annotations."],"supporting_citations":[{"why":"supplies the pretrained latent diffusion model whose geometric priors DMS fine-tunes for view synthesis.","marker":"[55]"},{"why":"supplies the recipe for repurposing the text-to-image model as an image-prompt generator, which DMS adapts to stereo views.","marker":"[33]"},{"why":"supplies the per-pixel minimum warping loss that DMS extends to multi-baseline source views.","marker":"[22]"},{"why":"the baseline stereo network used in the ablation studies.","marker":"[71]"},{"why":"an early unsupervised stereo matching method whose training and evaluation protocol DMS follows.","marker":"[95]"},{"why":"the simulator that provides ground-truth multi-baseline views for evaluating generated novel views.","marker":"[15]"},{"why":"one of the supervised stereo networks adapted to self-supervised fine-tuning and improved by DMS.","marker":"[80]"},{"why":"another stereo network used to demonstrate that DMS generalizes across learning-based matchers.","marker":"[39]"},{"why":"the geometry-based multi-baseline synthesis approach used as the comparison baseline for generation quality.","marker":"[23]"},{"why":"the stereo-supervised monocular depth estimation baseline that DMS is shown to improve.","marker":"[21]"}],"fun_headline_variants":["Diffusion-generated views slash depth outlier errors by 35%","Novel-view diffusion boosts self-supervised depth accuracy","Multi-baseline stereo from diffusion cuts depth outliers 35%","Diffusion model fills occlusions to sharpen depth estimates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that upscaling the diffusion model's input by a factor X shrinks the generated horizontal displacement by the same factor, so the X=2 output really is a center view; the paper treats this as an empirical finding and supports it only indirectly with PSNR/SSIM on CARLA, not by measuring the generated center view's disparity.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-generated views slash depth outlier errors by 35%","Novel-view diffusion boosts self-supervised depth accuracy","Multi-baseline stereo from diffusion cuts depth outliers 35%","Diffusion model fills occlusions to sharpen depth estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00039,"raw_usage":{"total_tokens":2042,"prompt_tokens":923,"completion_tokens":1119,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":1060}},"tokens_in":539,"tokens_out":1119,"duration_ms":9109,"temperature":1.0,"reasoning_tokens":1060,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:15:25.839972+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a CARLA scene with ground-truth multi-baseline views, generate the center view from the left image at X=2 and compute the end-point error between this generated view warped by true disparity and the true center image; if the center view does not sit at half the left-right disparity, the geometric-consistency assumption behind the minimum warping loss fails.","supporting_citations":[{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"supplies the pretrained latent diffusion model whose geometric priors DMS fine-tunes for view synthesis."},{"cited_title":"Repurpos- ing diffusion-based image generators for monocular depth estimation","cited_arxiv_id":null,"evidence_quote":"supplies the recipe for repurposing the text-to-image model as an image-prompt generator, which DMS adapts to stereo views."},{"cited_title":"Parallax attention for unsupervised stereo correspondence learning","cited_arxiv_id":null,"evidence_quote":"the baseline stereo network used in the ablation studies."},{"cited_title":"Un- supervised learning of stereo matching","cited_arxiv_id":null,"evidence_quote":"an early unsupervised stereo matching method whose training and evaluation protocol DMS follows."},{"cited_title":"Iterative geometry encoding volume for stereo matching","cited_arxiv_id":null,"evidence_quote":"one of the supervised stereo networks adapted to self-supervised fine-tuning and improved by DMS."},{"cited_title":"Raft-stereo: Multilevel recurrent field transforms for stereo matching","cited_arxiv_id":null,"evidence_quote":"another stereo network used to demonstrate that DMS generalizes across learning-based matchers."}],"review_version":2}