{"id":"3efa2b8b-d365-4bc4-92c9-4a623a04eb13","arxiv_id":"2505.05509","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"StereoINR achieves arbitrary-scale stereo super-resolution by fusing implicit neural representation with disparity-guided cross-view attention, improving stereo consistency over fixed-scale methods.","lead":"This paper presents StereoINR, a method that uses a continuous neural representation to upscale stereo image pairs at any zoom level while keeping the left and right views consistent. It combines implicit neural networks with stereo disparity information and cross-view attention, and reports better cross-view consistency than prior stereo super-resolution methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Disparity accuracy is the un-tested load-bearing premise: the SCORE metric uses a different disparity estimator than the SPyNet warping inside the model, and the paper's own limitation admits reliance on accurate disparity.","rationale":"Reader's conditional verdict identified SPyNet disparity accuracy as the weakest assumption; I agree that this is the load-bearing point, and the paper's own limitations section supports it. My concern sharpens it: the supporting metric (SCORE) is computed with a different disparity estimator (RAFT-Stereo) than the one used inside the architecture (SPyNet). Since no explicit geometric loss is imposed, and the only novelty relative to single-image INR is the disparity-guided cross-attention, the empirical case for 'pixel-level geometric consistency' reduces to whether SPyNet's disparity is good enough and whether SCORE is a faithful proxy. Neither is established. The proposed test is feasible because Middlebury provides GT disparity; KITTI also has semi-dense GT, enabling occlusion-split analysis. If the GT-disparity check preserves the SCORE advantage and inference is invariant to replacing SPyNet with GT, the concern is resolved and the central claim stands. If not, the paper should be revised to either add a disparity-robustness analysis, an explicit consistency loss, or soften the geometric-consistency claim. The missing ASteISR baseline and lack of error bars are secondary; they affect completeness, not the core argument. I therefore see no reason to change the reader's CONDITIONAL verdict.","tokens_in":13966,"tokens_out":8433,"duration_ms":90365,"concrete_test":"On Middlebury (which has ground-truth disparity), recompute Table 2's SCORE using GT disparity for L(D_SR,D_HR) instead of RAFT-Stereo estimates; if StereoINR's SCORE advantage over SwinFIRSSR/NAFSSR shrinks or reverses, the geometric-consistency claim is metric-dependent. As a second arm, run inference with GT disparity replacing SPyNet in Eqs. 5-8 at x2 and x4; if PSNR/SCORE change materially, the method's reliance on accurate disparity is confirmed, and the paper must quantify failure at occlusions and non-rectified inputs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of 'significant improvements in pixel-level geometric consistency' (abstract) rests on SCORE (Eq. 18), which estimates disparity from SR/HR outputs with RAFT-Stereo. Inside the model, however, alignment is driven by SPyNet optical-flow-as-disparity on LR inputs (Section 4.1, Eqs. 5-8). The paper never measures SPyNet's disparity error on its test sets, nor tests sensitivity to it; Section 5 lists 'relies on accurate disparity estimation' as a limitation but provides no experiment. Because warping with a wrong disparity misaligns keys/values in cross-attention (Eqs. 9-12), and because no explicit left-right consistency or disparity loss is optimized (training is L1 per view, Section 4.1), the geometry-consistency result is an emergent property whose magnitude is only as trustworthy as the SPyNet prior. At occlusions, textureless regions, and non-rectified pairs, the 'horizontal optical flow = disparity' assumption fails; these failure modes are not analyzed. In addition, SCORE's disparity term uses a different learned estimator (RAFT-Stereo), so the reported large SCORE gains could partly reflect alignment cues inherited from SPyNet rather than verifiable subpixel geometry. This is the load-bearing concern because Table 2's headline differentiator is SCORE, while PSNR/SSIM gains over SwinFIRSSR are marginal or negative.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes StereoINR, a method for arbitrary-scale stereo image super-resolution. It combines a frozen HAT-based encoder extended with stereo, scale, and spatial adapters, with a disparity-guided arbitrary-scale upsampler (DGASU). DGASU estimates disparity from the low-resolution left-right pair using SPyNet horizontal optical flow, warps deep features across views, and then applies alternating self- and cross-attention in an implicit neural representation framework to produce continuous coordinate-to-color mappings. The method is evaluated on Middlebury, Flickr1024, and KITTI datasets under both out-of-distribution scales (Table 1, scales ×2 to ×30) and in-distribution scales (Table 2, ×2 and ×4), with PSNR, SSIM, LPIPS, and a stereo consistency metric SCORE. Ablations study the adapters, the upsampler design, and the attention mechanism.","tokens_in":14294,"tokens_out":4141,"duration_ms":41616,"significance":"If the geometric-consistency improvements are trustworthy, StereoINR would be a useful contribution: it offers a single model for continuous magnification of stereo pairs while explicitly fusing cross-view information through disparity-guided warping. The paper contains multi-dataset and multi-scale comparisons, component-wise ablations, and an honest limitations section. The idea of warping features with estimated disparity before cross-attention is reasonable, and the out-of-scale comparisons in Table 1 show consistent gains over INR-based single-image super-resolution baselines. However, the paper provides no code, no error bars, and only 112 validation pairs, and the central geometric-consistency claim rests on an unvalidated disparity prior and a metric computed with a different estimator.","major_comments":[{"comment":"The disparity guidance is a load-bearing component, but its accuracy is never measured. The model uses SPyNet horizontal optical flow as stereo disparity on low-resolution inputs to warp features for cross-attention (Eqs. 7-12), yet the paper reports no disparity error on any test set, no occlusion analysis, and no sensitivity study. Because training uses only per-view L1 loss with no disparity-consistency term, the claimed geometric consistency is an emergent property whose reliability is unknown. I recommend adding an evaluation of SPyNet disparity error on a dataset with ground-truth disparity, or an ablation comparing SPyNet warping with RAFT-Stereo warping and with ground-truth disparity, and an analysis of behaviors at occlusions and textureless regions.","section":"Section 4.1, Eqs. (5)-(12)"},{"comment":"There is an estimator mismatch between the consistency metric and the model's internal alignment. SCORE (Eq. 18) estimates disparity from SR/HR outputs with RAFT-Stereo, while the model aligns features with SPyNet on LR inputs. The large SCORE improvements in Table 2 (e.g., Flickr1024 ×4 SCORE 0.5544 vs. SwinFIRSSR 0.3211) could partly reflect an alignment bias inherited from SPyNet rather than verifiable subpixel geometry. To support the claim of 'significant improvements in pixel-level geometric consistency,' the authors should also report SCORE using a disparity estimator unrelated to the architecture, and should decompose SCORE into its perceptual and disparity terms so that the reader can see how much of the gain comes from the disparity term.","section":"Eq. (18) and Table 2"},{"comment":"The claim of outperforming existing approaches is contradicted by the paper's own numbers. The introduction states that the method 'outperform[s] existing approaches in both qualitative and quantitative evaluations,' but Table 2 shows StereoINR does not beat SwinFIRSSR on PSNR or SSIM in most in-scale cells: for example, KITTI2012 ×2 PSNR 31.58 vs. 31.79 and SSIM 0.9289 vs. 0.9321, KITTI2012 ×4 PSNR 27.14 vs. 27.16, and Flickr1024 ×2 SSIM 0.9254 vs. 0.9286. The performance claim should be restricted to SCORE and, for PSNR/LPIPS, to the out-of-distribution scales in Table 1; the abstract's language about 'matching' state-of-the-art methods is more accurate and should be used consistently.","section":"Introduction, third contribution, and Table 2"},{"comment":"The quantitative conclusions are not supported with measures of variability. The validation set is only 112 pairs, and no error bars, confidence intervals, or significance tests are reported. Several PSNR differences are on the order of 0.01-0.03 dB (e.g., Middlebury ×4 30.47 vs. SwinFIRSSR 30.44), which is within typical run-to-run variation for a model trained with batch size 1 and no reported number of seeds. I ask for standard deviations across at least three training runs, or paired statistical tests on the validation set, for the main comparisons in Tables 1 and 2.","section":"Section 4.1 and Tables 1-2"}],"minor_comments":[{"comment":"There is an extra closing parenthesis in the equation: 'Fθ(ez*L, ez*R, γ(x−x∗)))' should be 'Fθ(ez*L, ez*R, γ(x−x∗))'.","section":"Eq. (17)"},{"comment":"The phrase 'the stereo adapter leverage cross-attention' should be 'leverages' for subject-verb agreement.","section":"Section 3.2"},{"comment":"The text cites reference [32] for RAFT-Stereo, but the reference is the RAFT optical flow paper; the RAFT-Stereo paper should be cited separately.","section":"References [32]"},{"comment":"The check marks and crosses in the table rows are not defined in the caption; adding a legend such as '✓ indicates that the adapter is enabled' would make the ablation results easier to interpret.","section":"Table 3"},{"comment":"The metric is introduced as 'Score' and used as 'SCORE' elsewhere; the capitalization should be made consistent.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is promising, but the headline claims depend on an unvalidated disparity prior and a small validation set. I would encourage the editor to request that the authors release code and trained models as part of the revision, since the empirical nature of the claims makes reproducibility essential. The SCORE-vs-PSNR tension should be resolved before the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: StereoINR is a well-built integration of INR upsampling, adapter-tuned HAT features, and disparity-warped cross-attention. The out-of-scale results are consistently ahead of the INR baselines, and the ablations support each design choice. But the headline claim about pixel-level geometry consistency rests on a single learned metric, and the paper never independently validates the SPyNet disparity that drives the whole cross-view fusion. That is a real soft spot, not a fatal one.\n\nWhat is new: applying LIIF-style implicit representation to stereo pairs, with a disparity-guided upsampler that warps features before cross-attention. That is a sensible and non-obvious combination, and the authors show it beats SwinIR-LTE and EQSR at out-of-training scales by a comfortable margin. The adapter tuning of HAT for stereo is also a reasonable way to bring in strong single-image priors, and it keeps the trainable parameter count low (2.03M vs. SwinFIRSSR's 24M).\n\nWhere I have concerns. First, the consistency metric. Internally, the model warps features using SPyNet optical flow as disparity. Externally, SCORE is computed with RAFT-Stereo. The paper never measures SPyNet's error on these datasets, never ablate with a better/worse disparity estimator, and does not train any left-right consistency loss. The authors admit the method relies on accurate disparity, but they offer no sensitivity analysis. The large SCORE gains (e.g., 0.5544 vs 0.3211 on Flickr1024 at ×4) could be genuine, but they could also reflect the model imprinting SPyNet's disparity into the outputs. A simple test would be to swap SPyNet for RAFT inside the model, or to compute SCORE with a third estimator. Until then, the consistency claim is credible but unproven.\n\nSecond, the conclusion says the method outperforms existing approaches, but Table 2 shows SwinFIRSSR is ahead on PSNR on all three datasets at ×2, and on two of three at ×4. The abstract's softer wording (\"matches state-of-the-art\") is fine; the conclusion is not. Third, ASteISR is cited as the inspiration for adapter-based transfer but never appears in any comparison table. That is a glaring omission given how close the method families are. Fourth, no code and no error bars; given the small validation set (112 pairs), the single-run numbers should be treated cautiously.\n\nWho this is for: stereo SR researchers, especially those pushing arbitrary-scale upsampling. It is worth a serious referee, but I would condition acceptance on code release and on an experiment that decouples the disparity estimator from the consistency metric.\n\nMy recommendation: send it to peer review, not desk reject, and make the disparity sensitivity analysis a required revision.","headline":"Solid engineering with a genuinely new combination, but the geometry-consistency claim leans on an unvalidated disparity prior and the conclusion overreaches.","tokens_in":14781,"tokens_out":2842,"would_cite":false,"duration_ms":30336,"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 single implicit neural network can super-resolve stereo pairs at any scale while keeping views geometrically consistent.","keywords":["stereo image super-resolution","arbitrary-scale super-resolution","implicit neural representation","cross-view geometric consistency","disparity-guided upsampling","cross-attention","stereo consistency","adapter fine-tuning"],"falsifier":"Run StereoINR on rectified stereo pairs at a fixed scale and deliberately corrupt the disparity maps (for example, shift the horizontal flow by several pixels or replace SPyNet with a constant disparity) before the warping stage, then compare SCORE and PSNR in occluded versus non-occluded regions; if the consistency scores remain high under corrupted disparity, the disparity guidance is not the load-bearing ingredient, whereas a sharp drop would confirm that accurate disparity is required.","tokens_in":13789,"feed_emoji":"🖼️","tokens_out":7872,"duration_ms":72582,"temperature":0.7,"pith_summary":"StereoINR is a stereo image super-resolution method that treats a left-right image pair as one continuous implicit neural representation instead of discrete pixels, so a single trained model can upscale both views by any real-valued factor rather than being locked to a preset scale. The paper's central claim is that the upsampling stage itself should be cross-view aware: by estimating disparity with SPyNet, warping each view's deep features into the other view, and fusing them through cross-attention, the model preserves geometric consistency between the reconstructed left and right images. On Middlebury, Flickr1024, and KITTI, the authors report that StereoINR outperforms existing arbitrary-scale super-resolution methods at scales beyond those seen in training and matches leading fixed-scale stereo super-resolution methods within trained scales, while improving the disparity-based consistency metric SCORE. If the claim holds, stereo super-resolution no longer needs a separate network per scale, and the reconstructed pair can be trusted for downstream geometry tasks such as depth estimation and 3D reconstruction.","feed_headline":"One network upsamples stereo pairs at any scale, views kept consistent","feed_subtitle":"Warp-then-attend upsampling shares evidence across left and right views, beating fixed-scale methods on geometric consistency.","key_machinery":"The load-bearing mechanism is the disparity-guided arbitrary-scale upsampler (DGASU). Given latent codes from both views, it estimates bidirectional disparity with SPyNet, warps each view's latent code into the other view using bilinear warping, and forms cross-attention in which the current view's code is the query and the warped other-view code supplies keys and values; a squeeze-and-excitation refinement and alternation of self- and cross-attention layers then feed an MLP $F_\\theta: \\mathbb{R}^{2+C} \\to \\mathbb{R}^3$ that predicts RGB from continuous coordinates and nearest latent codes, with sinusoidal positional encoding $\\gamma(x)$ and a bicubic LR skip connection. This machinery is what lets a single model upsample at arbitrary real-valued scales while still letting each reconstructed pixel borrow evidence from its matching pixel in the other view.","core_discovery":"The paper argues that earlier stereo super-resolution methods produce inconsistent views because they upsample left and right features independently with convolution-based modules such as pixel shuffle, so the upsampling operation never uses information from the other view. StereoINR replaces that operation with a disparity-guided arbitrary-scale upsampler: a frozen single-image super-resolution transformer is adapted by lightweight stereo, scale, and spatial adapters to extract left and right latent codes; SPyNet supplies bidirectional disparity; each view's latent code is warped into the other view's coordinate frame; alternating self-attention and cross-attention aggregate aligned features; and an MLP maps continuous query coordinates, conditioned on the nearest latent codes, to output RGB values with a bicubic skip connection. The authors report that this design improves pixel-level geometric consistency measured by SCORE, outperforms other arbitrary-scale methods out of the training scale distribution, and matches fixed-scale methods within it.","pith_inferences":["If disparity-guided upsampling is the true source of the consistency gain, the same warp-then-attend idea should extend to light-field and multi-camera super-resolution, where more than two views must satisfy epipolar constraints.","A testable extension would jointly fine-tune the disparity estimator with the super-resolution objective; the paper uses a frozen SPyNet, so its accuracy sets an upper bound on how well cross-view features can align in occluded or textureless regions.","The largest consistency gains appear on Flickr1024 at $\\times 4$, a dataset with strong parallax and varied texture; this suggests the method's advantage will be most visible in scenes with genuine depth variation and may shrink for near-planar scenes.","Because the MLP maps continuous coordinates to colors given latent codes, the same representation could in principle output disparity or depth at arbitrary resolution alongside color, though the paper does not explore this."],"forward_implications":["A single StereoINR model can change its upsampling factor at inference time over a continuous range (the paper trains with scales sampled from $U(1,4)$ and tests up to $\\times 30$), removing the need to retrain or swap architectures per scale.","Because cross-view fusion happens during upsampling, the super-resolved left and right views stay consistent enough that a pre-trained stereo matcher (RAFT-Stereo) produces disparity maps closer to those of the ground-truth high-resolution pair, reflected in higher SCORE.","The frozen-encoder-plus-adapters design means only a small number of parameters are tuned (2.03M reported), so the approach is a parameter-efficient route from single-image super-resolution pretraining to stereo super-resolution.","Out-of-distribution scales from $\\times 6$ to $\\times 30$ show consistent PSNR and LPIPS gains over other arbitrary-scale methods such as LIIF, LTE, LINF, OPESR, and EQSR, according to the reported tables.","The model trained on Middlebury and Flickr1024 also performs on KITTI 2012 and KITTI 2015, indicating the learned cross-view upsampling transfers to new stereo domains."],"supporting_citations":[{"why":"introduces the local implicit image function that StereoINR extends to stereo pairs for coordinate-to-color mapping.","marker":"[6]"},{"why":"provides the pretrained HAT transformer whose RHAG blocks are frozen and adapted to stereo features.","marker":"[5]"},{"why":"NAFSSR supplies the stereo cross-attention design the stereo adapter resembles and serves as the main fixed-scale baseline.","marker":"[8]"},{"why":"SPyNet is the pretrained optical flow network used to estimate the disparity maps for warping.","marker":"[26]"},{"why":"RAFT-Stereo is the pretrained stereo matcher used to estimate disparity from SR and HR pairs for the SCORE metric.","marker":"[32]"},{"why":"ASteISR establishes the strategy of adapting a single-image super-resolution pretrained model to stereo, which StereoINR adopts with adapter tuning.","marker":"[49]"},{"why":"provides the Flickr1024 dataset used for training and qualitative/quantitative evaluation.","marker":"[37]"},{"why":"iPASSR provides the training and validation data splits and is a stereo super-resolution baseline based on parallax attention.","marker":"[38]"}],"fun_headline_variants":["StereoINR upsamples both views at any scale, geometry kept consistent","Implicit neural stereo super-res breaks scale limits, aligns views","Arbitrary-scale stereo upsampling with cross-view fusion","One model does any-scale stereo super-resolution, views stay consistent","Warp-attend-upsample: any-scale stereo super-res, consistent views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire cross-view fusion rests on the assumption that horizontal optical flow computed by the pre-trained SPyNet equals true disparity under rectified settings; wherever disparity is wrong, such as occlusions, textureless areas, or non-rectified inputs, the warped features misalign and the cross-attention can propagate incorrect information between views.","fun_headline_variants_meta":{"raw":{"variants":["StereoINR upsamples both views at any scale, geometry kept consistent","Implicit neural stereo super-res breaks scale limits, aligns views","Arbitrary-scale stereo upsampling with cross-view fusion","One model does any-scale stereo super-resolution, views stay consistent","Warp-attend-upsample: any-scale stereo super-res, consistent views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000814,"raw_usage":{"total_tokens":3563,"prompt_tokens":936,"completion_tokens":2627,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2535}},"tokens_in":552,"tokens_out":2627,"duration_ms":17972,"temperature":1.0,"reasoning_tokens":2535,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:33:37.828464+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run StereoINR on rectified stereo pairs at a fixed scale and deliberately corrupt the disparity maps (for example, shift the horizontal flow by several pixels or replace SPyNet with a constant disparity) before the warping stage, then compare SCORE and PSNR in occluded versus non-occluded regions; if the consistency scores remain high under corrupted disparity, the disparity guidance is not the load-bearing ingredient, whereas a sharp drop would confirm that accurate disparity is required.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the local implicit image function that StereoINR extends to stereo pairs for coordinate-to-color mapping."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the pretrained HAT transformer whose RHAG blocks are frozen and adapted to stereo features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"NAFSSR supplies the stereo cross-attention design the stereo adapter resembles and serves as the main fixed-scale baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SPyNet is the pretrained optical flow network used to estimate the disparity maps for warping."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RAFT-Stereo is the pretrained stereo matcher used to estimate disparity from SR and HR pairs for the SCORE metric."},{"cited_title":"ASteISR: Adapting Single Image Super-resolution Pre-trained Model for Efficient Stereo Image Super-resolution","cited_arxiv_id":"2407.03598","evidence_quote":"ASteISR establishes the strategy of adapting a single-image super-resolution pretrained model to stereo, which StereoINR adopts with adapter tuning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the Flickr1024 dataset used for training and qualitative/quantitative evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"iPASSR provides the training and validation data splits and is a stereo super-resolution baseline based on parallax attention."}],"review_version":1}