Pith. sign in

REVIEW 5 major objections 5 minor 116 references

StereoFlow proposes that stereo matching should blend deterministic disparity regression with generative distribution modeling, anchoring the generative model on the regression's output so that two sampling steps suffice to resolve ambiguit

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 11:04 UTC pith:VFK452G5

load-bearing objection StereoFlow is a plausible, well-ablated system for stereo matching with strong benchmark numbers, but the zero-shot claim is confounded by the DepthAnythingV2 prior and the manuscript has enough paste-in and labeling problems that I would not trust the tables until they are cleaned up. the 5 major comments →

arxiv 2607.19986 v1 pith:VFK452G5 submitted 2026-07-22 cs.CV

STEREOFLOW: Progressive Stereo Matching with StereoDiT and Transition Flow Matching

Hao Wang , Haoran Geng , Xiaotong Yang , Jing Tang , Songlin Wei , Linlong Lang , Yeying Jin , Zheng Zhu
show 2 more authors
Zhaoxin Fan Biao Leng
This is my paper
classification cs.CV
keywords stereo matchinggenerative modelingflow matchingdiffusion transformerdisparity estimationzero-shot generalizationambiguous regionsprogressive cascade
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that standard stereo matching, which estimates a single disparity value per pixel, systematically fails in ambiguous areas because the true correspondence is a multimodal distribution, not a point. The authors propose keeping a fast deterministic matching network to produce a strong disparity prior, then using a diffusion-style generative model to refine it by sampling from the distribution anchored around that prior. The key mechanism is a Transition Flow Matching objective that starts from the noisy prior estimate rather than from random noise, so the generative model learns to move geometric structures and synthesize high-frequency details in as few as two steps. On standard benchmarks, the full system reports the best error rates, including under zero-shot transfer from synthetic to real data. If correct, this points to a way of getting the robustness of generative modeling in stereo without the usual sampling cost.

Core claim

The paper's central claim is that stereo matching should not be framed as deterministic regression alone; the mismatch between the unimodal prediction and the intrinsically multimodal correspondence distribution causes systematic errors in ill-posed regions and limits generalization. StereoFlow instantiates a complementary formulation: a two-stage cascade matching network computes multi-resolution disparity priors, and a pixel-space diffusion transformer (StereoDiT) explicitly parameterizes the distribution of plausible disparities around those priors. The load-bearing identity is the Transition Flow Matching path, which replaces isotropic Gaussian noise with the degraded disparity estimate,

What carries the argument

The central mechanism is Transition Flow Matching, an optimal-coupling flow matching objective defined by a linear interpolant and a linear noise schedule. Starting from the deterministic network's noisy degraded disparity instead of isotropic Gaussian noise, it forces the generative model to learn a directional displacement of geometric structures (the interpolant) plus stochastic perturbation for detail synthesis (the noise schedule), enabling two-step Euler sampling. Around this objective sits StereoDiT, a pixel diffusion transformer split into a geometric encoder (large-patch, aligned to monocular depth representations) that anchors low-frequency structure, and a lightweight stereo decod

Load-bearing premise

The zero-shot generalization story assumes that the pretrained monocular depth foundation model has not already seen the real evaluation benchmarks during its own large-scale pretraining; if it has, the reported transfer gains could come from that monocular prior rather than from the stereo generative pipeline.

What would settle it

Train the identical StereoFlow architecture but replace the pretrained depth foundation model's weights with randomly initialized (or held-out) features, or evaluate on a new, freshly captured stereo dataset not present in any pretraining corpus. If the zero-shot gains over the deterministic baseline vanish, the claimed generalization benefit is an artifact of the monocular prior rather than of the generative stereo formulation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the reported results hold, stereo matchers can be built from a deterministic regression backbone plus a compact generative refiner, rather than choosing between the two paradigms.
  • A two-step generative refinement schedule is sufficient for high-quality disparity maps, which makes diffusion-style stereo practical at about 0.4 seconds per frame on a modern GPU.
  • Zero-shot synthetic-to-real transfer improves substantially, suggesting that explicitly modeling local ambiguity is a major factor in cross-domain robustness.
  • The frequency-decoupled design (geometry encoder vs. detail decoder) indicates that separating low-frequency consistency from high-frequency detail is an effective inductive bias for disparity refinement.
  • The progressive cascade paper claims that using different cost volume types at different resolutions (rather than the same type) is a core architectural contribution, improving both consistency and detail with fewer GRU iterations (16 vs. 32).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A confound worth testing: the depth foundation model that provides the geometric prior was pretrained on large real-world data; if it already resembles the evaluation benchmarks, part of the zero-shot gain may come from monocular depth priors rather than from the stereo generative pipeline. A controlled test would train the same architecture with the depth backbone frozen versus randomly initializ
  • The same Transition Flow Matching objective could be applied to other dense prediction tasks with strong priors, such as optical flow, monocular depth refinement, or surface normal estimation, wherever the output is a spatially structured field with local ambiguity.
  • The method's reliance on a large pretrained monocular depth model means the marginal cost of adding StereoDiT is architecture-specific; one-step sampling or distillation of the two-step sampler would determine whether the generative gains survive at real-time latency.
  • Because the paper's baseline is a simplified version of a prior method (with its monocular depth scaling removed), the incremental gain may be partly due to the cascade and flow objective; the ablations suggest each added component improves results incrementally.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes StereoFlow, a prior-guided generative stereo matching framework that combines deterministic cascade matching with generative distribution modeling. It introduces three components: (i) a two-stage progressive cascade network using a low-resolution 4D Geometry Encoding Volume and a high-resolution 3D All-pairs Correlation Volume; (ii) StereoDiT, a pixel diffusion transformer with a frequency-decoupled geometric encoder and stereo decoder, aligned to DepthAnythingV2 features via a REPA loss; and (iii) Transition Flow Matching, a few-step flow-matching objective that interpolates between degraded and target disparities with a linear interpolant and a linear noise scheduler. The paper reports state-of-the-art results on Scene Flow, KITTI, ETH3D, and Middlebury in both zero-shot and fine-tuned settings, with two-step sampling.

Significance. If the empirical claims held, the framework would be a meaningful step toward handling ambiguous and ill-posed stereo regions, where deterministic regression tends to over-smooth. The combination of a deterministic prior with a compact generative model is well motivated, and the reported two-step flow matching with sub-second inference is practically attractive. The manuscript includes a useful ablation of the cascade, StereoDiT, and the flow-matching objective. However, the current presentation contains several load-bearing inconsistencies: the zero-shot protocol is confounded by a real-data pre-trained monocular backbone, the claimed SOTA status is contradicted in places by the paper's own tables, and key training parameters (σ_min, σ_max, the matching-loss weight) are missing or inconsistently reported. These issues must be resolved before the central claims can be accepted.

major comments (5)
  1. [Section IV.A, Eq. (4), Table III] The 'zero-shot generalization' claim is confounded. The full model aligns StereoDiT's geometric encoder to DepthAnythingV2-L through the REPA loss (Eq. 4), and Section IV.A states that DepthAnythingV2-L is adopted as the depth foundation model. DepthAnythingV2 is pre-trained on large-scale real imagery, as the paper's own Related Work notes. Therefore the protocol 'All models are trained on synthetic Scene Flow and evaluated directly on realistic datasets' (Fig. 3 caption) is not a clean synthetic-to-real test for the complete system, and the Table III gains over DEFOM-Stereo/MonSter may stem substantially from the real-data monocular prior rather than from the proposed cascade or TFM. Because the strongest baselines also use monocular priors, the relative ranking does not isolate the paper's contributions. Please provide an ablation in which the depth foundation model is removed or repl
  2. [Section IV.B, Table II] The text claims 'the proposed model obtains the top performance on the ETH3D benchmark' (Section IV.B.1), but Table II shows MonSter at ETH3D Bad1.0 (noc) = 0.46 and Bad1.0 (all) = 0.72, while StereoFlow is 0.58 and 0.69. On Bad1.0 (noc) MonSter is clearly better. Similarly, in the zero-shot Table III, BridgeDepth achieves ETH3D = 1.3 while StereoFlow is 1.7, so the SOTA claim on ETH3D is not supported by the paper's own tables. Please correct the text and state precisely on which benchmarks/columns StereoFlow actually ranks first.
  3. [Eq. (11) and text after Eq. (11)] The overall loss in Eq. (11) uses γ_match L_match, but the sentence immediately after says 'where γ_iter = 0.1 defines the weight coefficient for the matching loss.' This is internally inconsistent: γ_iter is already used in Eq. (9) for the iteration weighting (γ_iter = 0.9). Please rename the matching-loss weight to γ_match consistently throughout. Additionally, the TFM objective in Eq. (6) depends on σ_min and σ_max, but their numerical values are never reported in the implementation details or ablations; without these, the exact objective cannot be reproduced.
  4. [Table II caption and Section III (inserted block)] Table II is captioned 'SYNTHETIC PRE-TRAINING EVALUATION ON SCENE FLOW TEST SET' but contains ETH3D, Middlebury, and KITTI columns; this caption is clearly mislabeled. More seriously, Section III contains a large block beginning 'StereoFlow: CascadeProgressiveStereo Matching with Transition Flow Matching' followed by repeated DEFOM-Stereo text, Figure 3 (DEFOM-Stereo overview), Section 3.2 ('Combined Feature Extraction'), and multiple repeated figures and equations. This appears to be an unintended inclusion of another paper's content and must be removed; it makes the manuscript unusable in its current form and raises integrity concerns even if the underlying research is sound.
  5. [Section IV.B, Table II relative improvements] Several relative improvement claims in the text do not match Table II. For Middlebury, the text says StereoFlow surpasses DEFOM-Stereo by 6% on Bad2.0 (noc) and 6% on Bad2.0 (all); the table gives 2.39→2.30 (3.8% relative) and 5.02→4.82 (4.0% relative). For KITTI-2012, the text says 8% on Out-2 (noc) and 5% on Out-2 (all); the table gives 1.43→1.32 (7.7%) and 1.79→1.71 (4.5%). Please recompute and correct these percentages.
minor comments (5)
  1. [Various] Typos: 'stardand practice' (Section IV.A), 'benckmark' (Sections IV.B.2/IV.B.4), 'intrudoce' (Sections II.C/IV.D), 'arctitecture' (Table IX).
  2. [Section III.A / Figure numbering] Because of the inserted DEFOM-Stereo block, figure and section numbering is confusing: Figure 3 appears both as a zero-shot qualitative comparison and as the DEFOM-Stereo overview. After removing the inserted block, renumber all figures and adjust references.
  3. [Fig. 3 caption] The caption states 'All models are trained on synthetic Scene Flow and evaluated directly on the realistic datasets.' Given the use of DepthAnythingV2-L, this statement is misleading; at minimum it should be qualified to indicate that the geometric encoder is trained with REPA alignment to a real-data-pretrained backbone.
  4. [Eq. (5)] The uncertainty input U to the Stereo decoder is described in words as 'warping the right image to the left and calculate the uncertainty'; please define the computation formally and specify how it is obtained.
  5. [Table VI / Table VII] Table VI reports sampling times of 0.04 s and 0.07 s for StereoDiT1/8 and StereoDiT1/4, while Table VII reports total inference time 0.39 s; please clarify whether the sampling times are per module and how they combine with the cascade matching time.

Circularity Check

0 steps flagged

No significant circularity: benchmark claims are external evaluations, and the flow-matching objective is a standard interpolant; the zero-shot confound is an empirical threat, not a circular derivation.

full rationale

The paper's central claims are empirical: state-of-the-art results on Scene Flow, KITTI, ETH3D, and Middlebury are supported by benchmark tables, not by a derivation in which target numbers are inserted into equations. The Transition Flow Matching objective (Eqs. 6-8) is the standard flow-matching / stochastic-interpolant velocity regression with a linear interpolant and a linear noise schedule; no benchmark metric is encoded in the loss, so no 'prediction' is forced by construction. The REPA alignment loss (Eq. 4) and the matching losses (Eqs. 9-11) are ordinary auxiliary objectives. The paper's few self-citations (e.g., D3RoMa [77]) appear only in related-work critiques and are not load-bearing. The main threat to the zero-shot claim is the use of DepthAnythingV2-L, a depth foundation model pre-trained on large-scale real data; this makes Table III a confounded transfer test, but that is a correctness/interpretability concern, not a circular derivation. The manuscript also contains an apparent verbatim insertion of DEFOM-Stereo's methodology text, which is an attribution issue rather than circularity. The paper's own declared limitations (diffusion latency and two-stage training instability) likewise do not indicate circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The model relies primarily on standard empirical machine-learning assumptions; no physical constants or mathematical theorems are derived. The main hidden choices are the loss weights, the unspecified noise-schedule endpoints, and the reliance on a pre-trained depth foundation model.

free parameters (4)
  • σ_min / σ_max noise schedule endpoints = not specified
    Eq. (6) defines σ_t = t σ_min + (1-t) σ_max, but the endpoints are never given values; they control the stochastic perturbation and likely affect two-step sampling quality.
  • γ_match (Eq. 11, text says γ_iter) = 0.1 (as stated in text)
    Weight balancing matching and generative losses; hand-set.
  • γ_repa = 0.5
    REPA alignment loss weight; hand-set; Eq. (10).
  • γ_iter = 0.9
    Exponential weighting for iterative disparity updates; Eq. (9).
axioms (5)
  • domain assumption Disparity distribution is intrinsically multimodal in ambiguous regions and can be parameterized generatively.
    Introduction motivates diffusion; not proven with data.
  • domain assumption Anchoring the generative model at a deterministic disparity prior improves efficiency without loss of correctness.
    Core design choice; validated only via ablation.
  • ad hoc to paper Frequency decoupling (geometric encoder low-freq, stereo decoder high-freq) is a beneficial inductive bias.
    Design choice specific to this architecture; no theorem.
  • domain assumption DepthAnythingV2-L provides useful monocular depth features for stereo.
    External model adopted as-is; its pretraining distribution affects zero-shot claims.
  • domain assumption Benchmark ground truth and thresholds are reliable enough for sub-pixel comparisons.
    Standard CV assumption; no error bars given in the paper.
invented entities (2)
  • StereoDiT independent evidence
    purpose: Pixel diffusion transformer to model correspondence ambiguity
    Not a physical entity; an architecture whose benchmark performance is a falsifiable handle. It builds on DiT/PixDiT.
  • Transition Flow Matching independent evidence
    purpose: Few-step flow matching objective coupling deterministic prior to target disparity
    A training objective; its validity is testable via the reported sampling results, but it is closely related to stochastic interpolants.

pith-pipeline@v1.3.0-alltime-deepseek · 26635 in / 12330 out tokens · 117256 ms · 2026-08-01T11:04:48.570421+00:00 · methodology

0 comments
read the original abstract

Stereo matching is a fundamental task in 3D reconstruction. Despite remarkable advances, the prevailing paradigms formulate stereo matching as a deterministic regression problem, collapsing the multimodal distribution modeling into a single-point estimation. This formulation suffers from a regression-to-mean bias, frequently struggling with ambiguous regions. In contrast, we introduce a prior-guided generative framework that integrates deterministic matching regression and generative distribution modeling within a complementary formulation. Built upon this formulation, we introduce StereoFlow through three key components: (i) a two-stage progressive cascade matching network that progressively produces multi-resolution stereo conditions with complementary matching cues; (ii) a pixel diffusion transformer (termed StereoDiT) with a frequency-decoupled architecture for modeling correspondence ambiguity; (iii) a few-step flow matching objective (termed Transition Flow Matching) for efficient optimization. In summary, \textsc{\textbf{StereoFlow}} achieves strong geometric consistency and rich fine-grained details in ill-posed, discontinuous regions and under zero-shot generalization. Extensive experiments demonstrate that the proposed StereoFlow establishes multiple state-of-the-art results across benchmarks, including Scene Flow, KITTI, ETH3D, and Middlebury.

Figures

Figures reproduced from arXiv: 2607.19986 by Biao Leng, Haoran Geng, Hao Wang, Jing Tang, Linlong Lang, Songlin Wei, Xiaotong Yang, Yeying Jin, Zhaoxin Fan, Zheng Zhu.

Figure 1
Figure 1. Figure 1: Comparisons on KITTI-2012 [2], KITTI-2015 [4], ETH3D [5] and Middlebury [6] with the recent state-of-the-art counterparts. The proposed STEREOFLOW achieves comprehensive performance improvements across various benchmarks. (Zoom in for a better view.) prior deterministic regression Matching Network Diffusion Model distribution modeling Cost Volume stereo condition confidence propagation Input Output PL PL :… view at source ↗
Figure 2
Figure 2. Figure 2: The prior-guided generative stereo framework that integrates deterministic matching regression and generative distribution modeling within a complementary formulation, where the former provides disparity priors that substantially constrain the correspondence search space, and the latter ex￾plicitly parameterizes multimodal distributions anchored around these priors. (Zoom in for a better view.) ambiguity o… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparisons of zero-shot generalization with the baselines IGEV-Stereo [7] and DEFOM-Stereo [8] on KITTI-2012 [2], KITTI￾2015 [4], and ETH3D [5]. All models are trained on the synthetic Scene Flow [9] and evaluated directly on the realistic datasets. In contrast, the proposed STEREOFLOW achieves strong geometric consistency and rich fine-grained details in ambiguous regions such as occlusions, … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our proposed DEFOM-Ster btiifl ftd tt d p ure and context encodersNextwe inser Combine Context Encoder [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparisons on ETH3D and Middlebury with the baselines IGEV-Stereo [7] and DEFOM-Stereo [8]. All models are finetuned on the realistic ETH3D [5] and Middlebury [6] datasets. In contrast, the proposed STEREOFLOW achieves strong geometric consistency and fine-grained details across challenging ill-posed regions, including reflective surfaces, highlights, textureless, occluded and discontinuous ar… view at source ↗
Figure 6
Figure 6. Figure 6: Sampling Visualization of Transition Flow Matching. Starting directly from the noisy degraded disparities, with only two sampling steps, the proposed Transition Flow Matching enables the displacement of geometric structures and the synthesis of high-frequency details, with improvements in challenging ill-posed regions, such as reflective surfaces, textureless, occluded and discontinuous areas. (Zoom in for… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

116 extracted references · 1 canonical work pages

  1. [1]

    A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,

    D. Scharstein and R. Szeliski, “A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,”IJCV, vol. 47, no. 1-3, pp. 7–42, 2002

    2002

  2. [2]

    Are we ready for autonomous driving? the KITTI vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the KITTI vision benchmark suite,” inCVPR, 2012, pp. 3354– 3361

    2012

  3. [3]

    Stereo matching in time: 100+ FPS video stereo matching for extended reality,

    Z. Cheng, J. Yang, and H. Li, “Stereo matching in time: 100+ FPS video stereo matching for extended reality,” inWACV, 2024, pp. 8704– 8713

    2024

  4. [4]

    Object scene flow for autonomous vehicles,

    M. Menze and A. Geiger, “Object scene flow for autonomous vehicles,” inCVPR, 2015, pp. 3061–3070

    2015

  5. [5]

    A multi-view stereo benchmark with high-resolution images and multi-camera videos,

    T. Sch ¨ops, J. L. Sch ¨onberger, S. Galliani, T. Sattler, K. Schindler, M. Pollefeys, and A. Geiger, “A multi-view stereo benchmark with high-resolution images and multi-camera videos,” inCVPR, 2017, pp. 2538–2547

    2017

  6. [6]

    High-resolution stereo datasets with subpixel-accurate ground truth,

    D. Scharstein, H. Hirschm ¨uller, Y . Kitajima, G. Krathwohl, N. Nesic, X. Wang, and P. Westling, “High-resolution stereo datasets with subpixel-accurate ground truth,” inGCPR, X. Jiang, J. Hornegger, and R. Koch, Eds., vol. 8753, 2014, pp. 31–42

    2014

  7. [7]

    Iterative geometry encoding volume for stereo matching,

    G. Xu, X. Wang, X. Ding, and X. Yang, “Iterative geometry encoding volume for stereo matching,” inCVPR, 2023, pp. 21 919–21 928

    2023

  8. [8]

    Defom-stereo: Depth foundation model based stereo matching,

    H. Jiang, Z. Lou, L. Ding, R. Xu, M. Tan, W. Jiang, and R. Huang, “Defom-stereo: Depth foundation model based stereo matching,” in CVPR, 2025, pp. 21 857–21 867

    2025

  9. [9]

    A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation,

    N. Mayer, E. Ilg, P. H ¨ausser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox, “A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation,” inCVPR, 2016, pp. 4040–4048

    2016

  10. [10]

    On the synergies between machine learning and binocular stereo for depth estimation from images: A survey,

    M. Poggi, F. Tosi, K. Batsos, P. Mordohai, and S. Mattoccia, “On the synergies between machine learning and binocular stereo for depth estimation from images: A survey,”IEEE TPAMI, vol. 44, no. 9, pp. 5314–5334, 2022

    2022

  11. [11]

    A survey on deep stereo matching in the twenties,

    F. Tosi, L. Bartolomei, and M. Poggi, “A survey on deep stereo matching in the twenties,”CoRR, vol. abs/2407.07816, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Pyramid stereo matching network,

    J. Chang and Y . Chen, “Pyramid stereo matching network,” inCVPR, 2018, pp. 5410–5418

    2018

  13. [13]

    On the over-smoothing problem of CNN based disparity estimation,

    C. Chen, X. Chen, and H. Cheng, “On the over-smoothing problem of CNN based disparity estimation,” inICCV. IEEE, 2019

    2019

  14. [14]

    Adaptive multi-modal cross-entropy loss for stereo matching,

    P. Xu, Z. Xiang, C. Qiao, J. Fu, and T. Pu, “Adaptive multi-modal cross-entropy loss for stereo matching,” inCVPR, 2024, pp. 5135– 5144

    2024

  15. [15]

    Raft-stereo: Multilevel recurrent field transforms for stereo matching,

    L. Lipson, Z. Teed, and J. Deng, “Raft-stereo: Multilevel recurrent field transforms for stereo matching,” in3DV, 2021, pp. 218–227

    2021

  16. [16]

    Parallax attention for unsupervised stereo correspondence learning,

    L. Wang, Y . Guo, Y . Wang, Z. Liang, Z. Lin, J. Yang, and W. An, “Parallax attention for unsupervised stereo correspondence learning,” IEEE TPAMI, vol. 44, no. 4, pp. 2108–2125, 2022

    2022

  17. [17]

    Deep stereo using adaptive thin volume representation with uncertainty awareness,

    S. Cheng, Z. Xu, S. Zhu, Z. Li, L. E. Li, R. Ramamoorthi, and H. Su, “Deep stereo using adaptive thin volume representation with uncertainty awareness,” inCVPR, 2020, pp. 2521–2531

    2020

  18. [18]

    Uncertainty estimation for stereo matching based on evidential deep learning,

    C. Wang, X. Wang, J. Zhang, L. Zhang, X. Bai, X. Ning, J. Zhou, and E. R. Hancock, “Uncertainty estimation for stereo matching based on evidential deep learning,”Pattern Recognit., vol. 124, p. 108498, 2022

    2022

  19. [19]

    Elfnet: Evidential local- global fusion for stereo matching,

    J. Lou, W. Liu, Z. Chen, F. Liu, and J. Cheng, “Elfnet: Evidential local- global fusion for stereo matching,” inICCV, 2023, pp. 17 738–17 747

    2023

  20. [20]

    Latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction,

    C. Wewer, K. Raj, E. Ilg, B. Schiele, and J. E. Lenssen, “Latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction,” inComputer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXXVII, ser. Lecture Notes in Computer Science, A. Leonardis, E. Ricci, S. Roth, O. Rus...

    work page doi:10.1007/978-3-031-73021 2024

  21. [21]

    Diffusion model for dense matching,

    J. Nam, G. Lee, S. Kim, H. Kim, H. Cho, S. Kim, and S. Kim, “Diffusion model for dense matching,” inICLR, 2024

    2024

  22. [22]

    Diffsplat: Repurposing image diffusion models for scalable gaussian splat generation,

    C. Lin, P. Pan, B. Yang, Z. Li, and Y . Mu, “Diffsplat: Repurposing image diffusion models for scalable gaussian splat generation,” in The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025. [Online]. Available: https://openreview.net/forum?id=eajZpoQkGK

    2025

  23. [23]

    DDT: decoupled diffusion transformer,

    S. Wang, Z. Tian, W. Huang, and L. Wang, “DDT: decoupled diffusion transformer,”CoRR, vol. abs/2504.05741, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Deco: Frequency- decoupled pixel diffusion for end-to-end image generation,

    Z. Ma, L. Wei, S. Wang, S. Zhang, and Q. Tian, “Deco: Frequency- decoupled pixel diffusion for end-to-end image generation,”arXiv preprint arXiv:2511.19365, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Image super-resolution via iterative refinement,

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image super-resolution via iterative refinement,”IEEE TPAMI, vol. 45, no. 4, pp. 4713–4726, 2023

    2023

  26. [26]

    Resshift: Efficient diffusion model for image super-resolution by residual shifting,

    Z. Yue, J. Wang, and C. C. Loy, “Resshift: Efficient diffusion model for image super-resolution by residual shifting,” inNeurIPS, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., 2023

    2023

  27. [27]

    Group-wise correlation stereo network,

    X. Guo, K. Yang, W. Yang, X. Wang, and H. Li, “Group-wise correlation stereo network,” inCVPR, 2019, pp. 3273–3282

    2019

  28. [28]

    IGEV++: iterative multi-range geometry encoding volumes for stereo matching,

    G. Xu, X. Wang, Z. Zhang, J. Cheng, C. Liao, and X. Yang, “IGEV++: iterative multi-range geometry encoding volumes for stereo matching,” CoRR, vol. abs/2409.00638, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    RAFT: recurrent all-pairs field transforms for optical flow,

    Z. Teed and J. Deng, “RAFT: recurrent all-pairs field transforms for optical flow,” inECCV, ser. Lecture Notes in Computer Science, A. Vedaldi, H. Bischof, T. Brox, and J. Frahm, Eds., vol. 12347. Springer, 2020, pp. 402–419

    2020

  30. [30]

    High- frequency stereo matching network,

    H. Zhao, H. Zhou, Y . Zhang, J. Chen, Y . Yang, and Y . Zhao, “High- frequency stereo matching network,” inCVPR, 2023, pp. 1327–1336

    2023

  31. [31]

    Mocha-stereo: Motif channel attention network for stereo matching,

    Z. Chen, W. Long, H. Yao, Y . Zhang, B. Wang, Y . Qin, and J. Wu, “Mocha-stereo: Motif channel attention network for stereo matching,” inCVPR, 2024, pp. 27 768–27 777

    2024

  32. [32]

    Selective-stereo: Adaptive frequency information selection for stereo matching,

    X. Wang, G. Xu, H. Jia, and X. Yang, “Selective-stereo: Adaptive frequency information selection for stereo matching,” inCVPR, 2024, pp. 19 701–19 710

    2024

  33. [33]

    Depth anything V2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything V2,” inNeurIPS, A. Globersons, L. Mackey, D. Bel- grave, A. Fan, U. Paquet, J. M. Tomczak, and C. Zhang, Eds., 2024

    2024

  34. [34]

    End-to-end learning of geometry and context for deep stereo regression,

    A. Kendall, H. Martirosyan, S. Dasgupta, and P. Henry, “End-to-end learning of geometry and context for deep stereo regression,” inICCV, 2017, pp. 66–75

    2017

  35. [35]

    Pcw-net: Pyramid combination and warping cost volume for stereo matching,

    Z. Shen, Y . Dai, X. Song, Z. Rao, D. Zhou, and L. Zhang, “Pcw-net: Pyramid combination and warping cost volume for stereo matching,” inECCV, S. Avidan, G. J. Brostow, M. Ciss ´e, G. M. Farinella, and T. Hassner, Eds., vol. 13692, 2022, pp. 280–297

    2022

  36. [36]

    Attention concatenation volume for accurate and efficient stereo matching,

    G. Xu, J. Cheng, P. Guo, and X. Yang, “Attention concatenation volume for accurate and efficient stereo matching,” inCVPR, 2022, pp. 12 971– 12 980

    2022

  37. [37]

    Mobilestereonet: Towards lightweight deep networks for stereo matching,

    F. Shamsafar, S. Woerz, R. Rahim, and A. Zell, “Mobilestereonet: Towards lightweight deep networks for stereo matching,” inWACV. IEEE, 2022, pp. 677–686

    2022

  38. [38]

    Lightstereo: Channel boost is all you need for efficient 2d cost aggregation,

    X. Guo, C. Zhang, Y . Zhang, W. Zheng, D. Nie, M. Poggi, and L. Chen, “Lightstereo: Channel boost is all you need for efficient 2d cost aggregation,” inICRA. IEEE, 2025, pp. 8738–8744

    2025

  39. [39]

    Domain-invariant stereo matching networks,

    F. Zhang, X. Qi, R. Yang, V . Prisacariu, B. W. Wah, and P. H. S. Torr, “Domain-invariant stereo matching networks,” inECCV, ser. Lecture Notes in Computer Science, A. Vedaldi, H. Bischof, T. Brox, and J. Frahm, Eds., vol. 12347. Springer, 2020, pp. 420–439

    2020

  40. [40]

    S2m2: Scalable stereo matching model for reliable depth estimation,

    J. Min, Y . Jeon, J. Kim, and M. Choi, “S2m2: Scalable stereo matching model for reliable depth estimation,”CoRR, vol. abs/2507.13229, 2025

    arXiv 2025

  41. [41]

    Diving into the fusion of monocular priors for generalized stereo matching,

    C. Yao, L. Yu, Z. Liu, J. Zeng, Y . Wu, and Y . Jia, “Diving into the fusion of monocular priors for generalized stereo matching,”CoRR, vol. abs/2505.14414, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Revisiting stereo depth estimation from a sequence- to-sequence perspective with transformers,

    Z. Li, X. Liu, N. Drenkow, A. S. Ding, F. X. Creighton, R. H. Taylor, and M. Unberath, “Revisiting stereo depth estimation from a sequence- to-sequence perspective with transformers,” inICCV, 2021, pp. 6177– 6186

    2021

  43. [43]

    Context-enhanced stereo transformer,

    W. Guo, Z. Li, Y . Yang, Z. Wang, R. H. Taylor, M. Unberath, A. L. Yuille, and Y . Li, “Context-enhanced stereo transformer,” inECCV, ser. Lecture Notes in Computer Science, S. Avidan, G. J. Brostow, M. Ciss´e, G. M. Farinella, and T. Hassner, Eds., vol. 13692, 2022, pp. 263–279

    2022

  44. [44]

    Unifying flow, stereo and depth estimation,

    H. Xu, J. Zhang, J. Cai, H. Rezatofighi, F. Yu, D. Tao, and A. Geiger, “Unifying flow, stereo and depth estimation,”IEEE TPAMI, vol. 45, no. 11, pp. 13 941–13 958, 2023

    2023

  45. [45]

    Global occlusion-aware transformer for robust stereo matching,

    Z. Liu, Y . Li, and M. Okutomi, “Global occlusion-aware transformer for robust stereo matching,” inWACV. IEEE, 2024, pp. 3523–3532. SUBMISSION TO IEEE TRANSACTIONS 12

    2024

  46. [46]

    Cascade cost volume for high-resolution multi-view stereo and stereo matching,

    X. Gu, Z. Fan, S. Zhu, Z. Dai, F. Tan, and P. Tan, “Cascade cost volume for high-resolution multi-view stereo and stereo matching,” inCVPR, 2020, pp. 2492–2501

    2020

  47. [47]

    Cfnet: Cascade and fused cost volume for robust stereo matching,

    Z. Shen, Y . Dai, and Z. Rao, “Cfnet: Cascade and fused cost volume for robust stereo matching,” inCVPR, 2021, pp. 13 906–13 915

    2021

  48. [48]

    Practical stereo matching via cascaded recurrent network with adaptive correlation,

    J. Li, P. Wang, P. Xiong, T. Cai, Z. Yan, L. Yang, J. Liu, H. Fan, and S. Liu, “Practical stereo matching via cascaded recurrent network with adaptive correlation,” inCVPR, 2022, pp. 16 242–16 251

    2022

  49. [49]

    Uncertainty guided adaptive warping for robust and efficient stereo matching,

    J. Jing, J. Li, P. Xiong, J. Liu, S. Liu, Y . Guo, X. Deng, M. Xu, L. Jiang, and L. Sigal, “Uncertainty guided adaptive warping for robust and efficient stereo matching,” inICCV, 2023, pp. 3295–3304

    2023

  50. [50]

    Learning to adapt for stereo,

    A. Tonioni, O. Rahnama, T. Joy, L. D. Stefano, T. Ajanthan, and P. H. S. Torr, “Learning to adapt for stereo,” inCVPR, 2019, pp. 9661–9670

    2019

  51. [51]

    Matching-space stereo networks for cross-domain generalization,

    C. Cai, M. Poggi, S. Mattoccia, and P. Mordohai, “Matching-space stereo networks for cross-domain generalization,” in3DV, V . Struc and F. G. Fern ´andez, Eds. IEEE, 2020, pp. 364–373

    2020

  52. [52]

    Edgestereo: An effective multi-task learning network for stereo matching and edge detection,

    X. Song, X. Zhao, L. Fang, H. Hu, and Y . Yu, “Edgestereo: An effective multi-task learning network for stereo matching and edge detection,” IJCV, vol. 128, no. 4, pp. 910–930, 2020

    2020

  53. [53]

    Segstereo: Exploiting semantic information for disparity estimation,

    G. Yang, H. Zhao, J. Shi, Z. Deng, and J. Jia, “Segstereo: Exploiting semantic information for disparity estimation,” inECCV, ser. Lecture Notes in Computer Science, V . Ferrari, M. Hebert, C. Sminchisescu, and Y . Weiss, Eds., vol. 11211. Springer, 2018, pp. 660–676

    2018

  54. [54]

    Croco: Self- supervised pre-training for 3d vision tasks by cross-view completion,

    P. Weinzaepfel, V . Leroy, T. Lucas, R. Br ´egier, Y . Cabon, V . Arora, L. Antsfeld, B. Chidlovskii, G. Csurka, and J. Revaud, “Croco: Self- supervised pre-training for 3d vision tasks by cross-view completion,” inNeurIPS, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., 2022

    2022

  55. [55]

    Robust synthetic-to-real transfer for stereo matching,

    J. Zhang, J. Li, L. Huang, X. Yu, L. Gu, J. Zheng, and X. Bai, “Robust synthetic-to-real transfer for stereo matching,” inCVPR. IEEE, 2024, pp. 20 247–20 257

    2024

  56. [56]

    Revisiting domain generalized stereo matching networks from a feature consistency perspective,

    J. Zhang, X. Wang, X. Bai, C. Wang, L. Huang, Y . Chen, L. Gu, J. Zhou, T. Harada, and E. R. Hancock, “Revisiting domain generalized stereo matching networks from a feature consistency perspective,” in CVPR. IEEE, 2022, pp. 12 991–13 001

    2022

  57. [57]

    Learning representa- tions from foundation models for domain generalized stereo matching,

    Y . Zhang, L. Wang, K. Li, Y . Wang, and Y . Guo, “Learning representa- tions from foundation models for domain generalized stereo matching,” inECCV, ser. Lecture Notes in Computer Science, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds., vol. 15100. Springer, 2024, pp. 146–162

    2024

  58. [58]

    Foundationstereo: Zero-shot stereo matching,

    B. Wen, M. Trepte, J. Aribido, J. Kautz, O. Gallo, and S. Birchfield, “Foundationstereo: Zero-shot stereo matching,” inCVPR, 2025, pp. 5249–5260

    2025

  59. [59]

    Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail,

    L. Bartolomei, F. Tosi, M. Poggi, and S. Mattoccia, “Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail,” inCVPR, 2025, pp. 1013–1027

    2025

  60. [60]

    Monster: Marry monodepth to stereo unleashes power,

    J. Cheng, L. Liu, G. Xu, X. Wang, Z. Zhang, Y . Deng, J. Zang, Y . Chen, Z. Cai, and X. Yang, “Monster: Marry monodepth to stereo unleashes power,” inCVPR, 2025, pp. 6273–6282

    2025

  61. [61]

    Bridgedepth: Bridging monocular and stereo reasoning with latent alignment,

    T. Guan, J. Guo, C. Wang, and Y . Liu, “Bridgedepth: Bridging monocular and stereo reasoning with latent alignment,”CoRR, vol. abs/2508.04611, 2025

    Pith/arXiv arXiv 2025

  62. [62]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” inICLR. OpenReview.net, 2021

    2021

  63. [63]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inICLR. OpenReview.net, 2023

    2023

  64. [64]

    Stochastic interpolants: A unifying framework for flows and diffusions,

    M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden, “Stochastic interpolants: A unifying framework for flows and diffusions,”CoRR, vol. abs/2303.08797, 2023

    Pith/arXiv arXiv 2023

  65. [65]

    Depthfm: Fast generative monocular depth estimation with flow matching,

    M. Gui, J. Schusterbauer, U. Prestel, P. Ma, D. Kotovenko, O. Grebenkova, S. A. Baumann, V . T. Hu, and B. Ommer, “Depthfm: Fast generative monocular depth estimation with flow matching,” in AAAI, T. Walsh, J. Shah, and Z. Kolter, Eds., 2025, pp. 3203–3211

    2025

  66. [66]

    Lotus: Diffusion-based visual foundation model for high- quality dense prediction,

    J. He, H. Li, W. Yin, Y . Liang, L. Li, K. Zhou, H. Zhang, B. Liu, and Y . Chen, “Lotus: Diffusion-based visual foundation model for high- quality dense prediction,” inICLR. OpenReview.net, 2025

    2025

  67. [67]

    Pixel- perfect depth with semantics-prompted diffusion transformers,

    G. Xu, H. Lin, H. Luo, X. Wang, J. Yao, L. Zhu, Y . Pu, C. Chi, H. Sun, B. Wang, G. Chen, H. Ye, S. Peng, and X. Yang, “Pixel- perfect depth with semantics-prompted diffusion transformers,”CoRR, vol. abs/2510.07316, 2025

    arXiv 2025

  68. [68]

    MVDD: multi-view depth diffusion mod- els,

    Z. Wang, Q. Xu, F. Tan, M. Chai, S. Liu, R. Pandey, S. Fanello, A. Kadambi, and Y . Zhang, “MVDD: multi-view depth diffusion mod- els,” inECCV, ser. Lecture Notes in Computer Science, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds., vol. 15071. Springer, 2024, pp. 236–253

    2024

  69. [69]

    Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction,

    H. Chen, J. Gu, A. Chen, W. Tian, Z. Tu, L. Liu, and H. Su, “Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction,” inICCV, 2023, pp. 2416–2425

    2023

  70. [70]

    Deterministic point cloud diffusion for denoising,

    Z. Liu, Z. Huang, M. Pan, and Y . He, “Deterministic point cloud diffusion for denoising,”IEEE Transactions on Visualization and Computer Graphics, vol. 32, no. 2, pp. 1822–1834, 2026

    2026

  71. [71]

    Exploiting diffusion prior for real-world image super-resolution,

    J. Wang, Z. Yue, S. Zhou, K. C. K. Chan, and C. C. Loy, “Exploiting diffusion prior for real-world image super-resolution,”International Journal of Computer Vision, pp. 1–21, 2024

    2024

  72. [72]

    Residual denoising diffusion models,

    J. Liu, Q. Wang, H. Fan, Y . Wang, Y . Tang, and L. Qu, “Residual denoising diffusion models,” inCVPR, June 2024, pp. 2773–2783

    2024

  73. [73]

    Diffcap: Diffusion-based real-time human motion capture using sparse imus and a monocular camera,

    S. Pan, X. Yi, Y . Zhou, W. Jian, Y . Zhang, P. Wan, and F. Xu, “Diffcap: Diffusion-based real-time human motion capture using sparse imus and a monocular camera,”IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 12, pp. 10 272–10 283, 2025

    2025

  74. [74]

    Coshmdm: Contact and shape-aware latent motion diffusion model for human in- teraction generation,

    A. A. Manjotho, T. T. Tewolde, R. A. Duma, and Z. Niu, “Coshmdm: Contact and shape-aware latent motion diffusion model for human in- teraction generation,”IEEE Transactions on Visualization & Computer Graphics, no. 01, pp. 1–14, Mar. 5555

  75. [75]

    Coreeditor: Correspondence- constrained diffusion for consistent 3d editing,

    Z. Zhu, H. Chen, P. Li, and M. Wei, “Coreeditor: Correspondence- constrained diffusion for consistent 3d editing,”IEEE Transactions on Visualization & Computer Graphics, vol. 32, no. 03, pp. 2838–2851, Mar. 2026. [Online]. Available: https://doi.ieeecomputersociety.org/10. 1109/TVCG.2026.3657658

    arXiv 2026

  76. [76]

    Diffuvolume: Diffusion model for volume based stereo matching,

    D. Zheng, X. Wu, Z. Liu, J. Meng, and W. Zheng, “Diffuvolume: Diffusion model for volume based stereo matching,”Int. J. Comput. Vis., vol. 133, no. 7, pp. 3807–3821, 2025

    2025

  77. [77]

    D3roma: Disparity diffusion-based depth sens- ing for material-agnostic robotic manipulation,

    S. Wei, H. Geng, J. Chen, C. Deng, W. Cui, C. Zhao, X. Fang, L. J. Guibas, and H. Wang, “D3roma: Disparity diffusion-based depth sens- ing for material-agnostic robotic manipulation,” inConference on Robot Learning, 6-9 November 2024, Munich, Germany, ser. Proceedings of Machine Learning Research, P. Agrawal, O. Kroemer, and W. Burgard, Eds., vol. 270. PM...

    2024

  78. [78]

    Flowdiffuser: Advancing optical flow estimation with diffusion models,

    A. Luo, X. Li, F. Yang, J. Liu, H. Fan, and S. Liu, “Flowdiffuser: Advancing optical flow estimation with diffusion models,” inCVPR. IEEE, 2024, pp. 19 167–19 176

    2024

  79. [79]

    Lightweight and accurate multi-view stereo with confidence-aware diffusion model,

    F. Wang, Q. Xu, Y . Ong, and M. Pollefeys, “Lightweight and accurate multi-view stereo with confidence-aware diffusion model,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 11, pp. 10 060–10 075, 2025

    2025

  80. [80]

    Rethinking iterative stereo matching from a diffusion bridge model perspective,

    Y . Shi, S. Zhou, W. Wang, and X. Lu, “Rethinking iterative stereo matching from a diffusion bridge model perspective,”Pattern Recognit., vol. 167, p. 111737, 2025

    2025

Showing first 80 references.