Pith. sign in

REVIEW 3 major objections 6 minor 51 references

Optical flow can be estimated faster and more accurately by treating hard motion as soft inpainting guided by a confidence mask.

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

2026-07-14 13:59 UTC pith:CED3AODT

load-bearing objection Solid hybrid engineering: soft confidence-guided inpainting of flow that beats FlowDiffuser on hard splits and cuts diffusion iterations a lot, with clean ablations and no load-bearing flaw. the 3 major comments →

arxiv 2607.10140 v1 pith:CED3AODT submitted 2026-07-11 cs.CV cs.AI

FlowPainter: Inpainting Optical Flow via Confidence-Guided Completion

Yuang Meng , Chenyang Wu , Xianshun Liu , Chun-Le Guo , Zichen Liang , Lina Lei , Jie Liang , Hui Zeng
show 2 more authors
Chongyi Li Lei Zhang
This is my paper
classification cs.CV cs.AI
keywords optical flowdiffusion modelsconfidence-guided inpaintingsoft residual guidancesimple-flow priorSintelKITTI
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.

Most of a scene's motion is easy for a small network to estimate; the hard parts are large displacements, occlusions, and messy motion. Pure diffusion models still denoise every pixel from pure noise, wasting effort on the easy parts and slowing training. FlowPainter first runs a lightweight confidence-aware network that produces a rough flow plus a per-pixel reliability mask, then treats the diffusion step as soft inpainting: high-confidence regions seed the process and supply residual guidance that decays over time, while low-confidence regions stay free for generative refinement. On standard benchmarks the hybrid reaches stronger accuracy than earlier diffusion flow methods under comparable data, and the diffusion stage needs far fewer training iterations. The practical payoff is a way to keep the generative strength of diffusion where it is needed without paying its full cost everywhere.

Core claim

FlowPainter shows that dense optical-flow generation can be reformulated as confidence-guided soft inpainting: a lightweight network supplies a simple-flow prior and reliability mask that initialize and residual-guide the diffusion process with time-decaying strength, so the model stabilizes early structure in reliable regions while retaining late-stage freedom in hard regions, yielding higher accuracy and substantially fewer diffusion training iterations than prior full-field diffusion flow estimators under comparable settings.

What carries the argument

Confidence-guided soft inpainting: the simple-flow prior Fs = Fr ⊙ Mconf is used for confidence-based initialization of the noisy state and for time-decayed residual guidance Rt injected into the denoiser features, with guidance weight Wt = (1 − tcur/T)·s that starts strong and relaxes.

Load-bearing premise

The confidence mask built from rough-flow error and occlusion checks must correctly mark reliable simple regions so the prior helps early denoising without systematically misleading the model in hard regions.

What would settle it

If, on the same training protocol, removing the confidence-based initialization and residual guidance (or replacing the supervised mask with pure noise) yields equal or better endpoint error and Fl-all on Sintel final and KITTI while still using only the reduced iteration budget, the claim that the confidence prior is what improves accuracy and convergence collapses.

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

Share X Bluesky LinkedIn Reddit HN

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

3 major / 6 minor

Summary. FlowPainter reformulates dense optical-flow estimation as confidence-guided soft inpainting. A lightweight PWC-Net-style Confidence-Aware Network predicts a rough flow Fr and a pixel-wise confidence mask Mconf (supervised by normalized MSE to ground truth plus forward-backward occlusion, Eqs. 1–2). The simple-flow prior Fs = Fr ⊙ Mconf is used for confidence-based initialization of the diffusion state (Eq. 5, τ = 0.5) and for time-decayed residual guidance (Eqs. 3, 6–7) inside a FlowDiffuser backbone. Under matched training data (no large-scale rigid-flow pretraining), the method reports 0.87/1.32 EPE on Sintel train clean/final, 1.01/1.71 on Sintel test, 3.02 Fl-all on KITTI-15 test, and 0.40 EPE on Spring, while reducing diffusion-stage iterations from 100k/200k/180k/50k to 20k/50k/50k/15k (and wall-clock from ~7d to ~2d on 8×A100). Ablations (Tabs. 4–6) support the mask construction, time-decay, inpainting, and threshold choices.

Significance. If the reported gains hold under independent reimplementation, the paper supplies a practical and well-motivated hybrid of discriminative priors with diffusion refinement for optical flow. The central engineering insight—that full-from-noise denoising wastes capacity on easy regions and that a reliability-gated prior can stabilize early steps without hard clamping—is clearly articulated (Sec. 3.1, Fig. 1–2) and is backed by matched comparisons to FlowDiffuser, public-benchmark numbers, and targeted ablations. Public code is promised. The work is incremental relative to FlowDiffuser and RAFT-style pipelines, but the efficiency and hard-split accuracy improvements are of genuine interest to the optical-flow community.

major comments (3)
  1. Sec. 4.1 and the efficiency claim: the paper states that FlowPainter converges with 20k/50k/50k/15k iterations versus FlowDiffuser’s 100k/200k/180k/50k and reduces wall-clock from 7d2h to 2d1h on 8×A100. Iteration counts alone do not establish that the prior is the sole cause of faster convergence, because the Confidence-Aware Network is itself pretrained and the two-stage schedule differs. A controlled experiment that freezes the same backbone and training budget while toggling only the confidence-guided initialization/guidance (or reports FLOPs/wall-clock per iteration under identical logging) is needed to make the efficiency claim load-bearing rather than correlational.
  2. Sec. 3.1–3.2 and Tab. 1: flow magnitude is used as the practical proxy for “hard” regions when constructing Mhf (Eq. 1) and when motivating the hybrid design. Tab. 1 only thresholds on ground-truth magnitude; it does not isolate occlusion, textureless, or motion-boundary subsets. Because Mgt = min(Mhf, Mocc) already mixes magnitude error with occlusion, residual bias of the prior into non-magnitude hard regions remains possible. A short breakdown of EPE/Fl-all on occlusion and motion-boundary masks (standard Sintel/KITTI splits) would confirm that high-confidence priors do not systematically harm those regions.
  3. Tables 2–3: all headline numbers appear to be single-run. Given that diffusion training is stochastic and the gains on Sintel clean are small (0.88→0.87 train; 1.02→1.01 test), multi-seed means and standard deviations (or at least two additional seeds for the main FlowPainter vs FlowDiffuser comparison) are required before the “stronger accuracy under comparable settings” claim can be treated as robust.
minor comments (6)
  1. Fig. 1 caption is truncated mid-sentence (“Our ...”); complete or rewrite for readability.
  2. Eq. (3): Wt = (1.0 − tcur/T)·s; clarify whether tcur runs from T down to 0 or 0 to T, and whether the schedule is applied identically in training (single random t) and inference (full trajectory).
  3. Sec. 3.2: the erosion kernel size (7×7) and the Norm operator in Eq. (1) are stated without sensitivity analysis; a one-line note on robustness would help.
  4. Table 2 Spring column mixes methods trained with and without Spring; a clearer “extra data” flag or separate block would avoid confusion.
  5. Fig. 4: unsupervised vs supervised confidence maps are shown but not quantified (e.g., IoU with Mgt); a small numeric inset would strengthen the claim that explicit supervision is necessary.
  6. Typos / style: “amount of computational cost” (Fig. 1 caption), inconsistent hyphenation of “time-step” / “time step”, and “Fl-all” vs “Fl-all” capitalization across tables.

Circularity Check

0 steps flagged

No significant circularity: empirical hybrid architecture evaluated on external public benchmarks with no self-definitional reductions or load-bearing self-citations.

full rationale

FlowPainter is a standard computer-vision method paper. Its central claims (confidence-guided soft inpainting of optical flow yields higher accuracy and fewer diffusion iterations than full-from-noise baselines under comparable training) are supported by training a Confidence-Aware Network (PWC-Net-style) whose mask is supervised by ground-truth flow MSE plus forward-backward occlusion (Eqs. 1-2), then using that prior for initialization and residual guidance inside a FlowDiffuser backbone. All reported EPE/Fl-all numbers are measured on held-out public splits of Sintel, KITTI and Spring that are independent of the model parameters. Ablations (Tabs. 4-6) vary the mask construction, time-decay and threshold τ=0.5; none of these quantities is defined in terms of the final metric. The backbone citation (FlowDiffuser) is external prior work, not a self-citation chain. There is no uniqueness theorem, fitted constant renamed as prediction, or ansatz smuggled via overlapping authors. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 2 invented entities

As an empirical CV methods paper the load-bearing content is architectural choices, hand-set hyperparameters, and the domain assumption that a lightweight estimator’s confidence map is a useful reliability gate for diffusion. No new physical entities are postulated; free parameters are the usual training and guidance knobs.

free parameters (5)
  • confidence initialization threshold τ = 0.5
    Binary mask B = 1[M_conf > τ] used for diffusion initialization; chosen by ablation (Tab. 6) with best value 0.5.
  • guidance scale s = 0.95
    Upper limit of simple-flow guidance strength in the time-decay schedule Wt = (1 − t_cur/T)·s; set slightly below 1 to keep the prior soft.
  • diffusion steps T = 6
    Number of sampling/training timesteps; inherited from FlowDiffuser and fixed for both train and test.
  • confidence-mask loss weight = 0.8
    L1 weight on M_conf during second-stage training of the Confidence-Aware Network.
  • flow / warping loss weights (stage 1) = 0.1 / 0.01
    EPE weight 0.1 and self-supervised warping weight 0.01 used to pre-train the rough-flow head before mask supervision.
axioms (3)
  • domain assumption A lightweight PWC-Net-style network can produce a rough flow whose per-pixel reliability is well approximated by a combination of prediction error and forward-backward occlusion.
    Stated in Sec. 3.1–3.2 and used to construct M_gt (Eqs. 1–2); the entire guidance pipeline rests on this reliability gate being informative.
  • ad hoc to paper Flow magnitude is a practical proxy for large-displacement difficulty when partitioning simple vs hard regions.
    Explicitly adopted in Sec. 3.1; other factors (texture, non-rigidity) are acknowledged but not modeled in the mask construction.
  • domain assumption Standard conditional diffusion training and sampling (as in FlowDiffuser) remain valid when residual guidance is additively injected into denoiser features.
    Assumed throughout Sec. 3.3; no new theoretical guarantee is offered for the modified reverse process.
invented entities (2)
  • confidence-gated residual guidance with time-decay schedule no independent evidence
    purpose: Injects the simple-flow prior softly into each denoising step while allowing late-stage flexibility.
    Core algorithmic object introduced in Sec. 3.3 (Eqs. 3–7); no independent existence outside this pipeline.
  • Confidence-Aware Network (rough flow + pixel-wise confidence mask) no independent evidence
    purpose: Supplies the simple-flow prior and reliability gate for soft inpainting.
    Lightweight module defined in Sec. 3.2; trained with a custom M_gt that mixes error and occlusion.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FlowPainter: Inpainting Optical Flow via Confidence-Guided Completion." pith.science (2026). https://pith.science/paper/CED3AODT

@misc{pith2026260710140,
  author       = {Pith},
  title        = {Pith review of: FlowPainter: Inpainting Optical Flow via Confidence-Guided Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CED3AODT}},
  note         = {Machine review of arXiv:2607.10140}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing optical flow methods broadly follow two paradigms: iterative optimization and diffusion-based estimation. Iterative methods, exemplified by RAFT, achieve high accuracy through recurrent refinement, but remain challenged by large displacements and complex motion. Diffusion-based methods introduce generative modeling and show promise in such ambiguous regions. However, existing diffusion models usually denoise the entire dense flow field from Gaussian noise, including simple regions where reliable motion can already be estimated by a lightweight network. This increases the denoising burden and may cause slow convergence and unstable training. To address this issue, we introduce FlowPainter, a diffusion-based optical flow framework that reformulates dense-flow generation as confidence-guided soft inpainting. FlowPainter employs a lightweight confidence-aware network to predict a rough flow and a pixel-wise confidence mask, distinguishing reliable simple regions from uncertain hard regions. The resulting simple-flow prior is used for confidence-based initialization and further injected into iterative denoising through confidence-gated residual guidance. With dynamically decaying guidance strength, FlowPainter stabilizes early denoising while preserving the flexibility of the diffusion model for late-stage detail refinement. Extensive experiments on public benchmarks, including Sintel, KITTI, and Spring, show that FlowPainter achieves strong accuracy under comparable training settings and converges more efficiently than existing diffusion-based optical flow methods, with notable gains on challenging benchmark splits. Our approach offers a practical way to integrate reliable discriminative priors with diffusion-based refinement for optical flow estimation. Our code is publicly available at https://github.com/mya012/FlowPainter.

Figures

Figures reproduced from arXiv: 2607.10140 by Chenyang Wu, Chongyi Li, Chun-Le Guo, Hui Zeng, Jie Liang, Lei Zhang, Lina Lei, Xianshun Liu, Yuang Meng, Zichen Liang.

Figure 2
Figure 2. Figure 2: An overview of previous diffusion-based methods and our proposed framework FlowPainter. Previous diffusion-based methods generated optical flow entirely from random Gaussian noise. FlowPainter first employs a vanilla network to generate simple optical flow for small motion regions, then uses this as prior information to guide the inpainting of the final optical flow. Based on this idea, we propose FlowPain… view at source ↗
Figure 3
Figure 3. Figure 3: An overview of our proposed framework FlowPainter, which consists of two main components: (a) Confidence-Aware Network: A lightweight PWC-Net-based model [35, 51] predicts a rough flow and a pixel-wise confidence mask indicating motion reliability. The confidence mask serves as a reliability gate, providing a simple-flow prior in reliable regions and suppressing unreliable predictions in challenging areas.… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of output results when training Confidence-Aware Network with or without explicit supervision of the confidence mask, where white denotes high confidence and black denotes low confidence. Training Details. 1) Training of Confidence-Aware Network. We train the Confidence-Aware Network in two stages. In the first stage, we pre-train on FlyingChairs [7] and FlyingThings3D [20], using supervision on… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of FlowPainter with SEA-RAFT [41] and FlowDiffuser [17] on Sintel(test). 4.3 Ablation Studies Confidence Mask Construction Strategy. The reliability of the confidence mask predicted by the Confidence-Aware Network is crucial for downstream inpainting, as it determines where the diffusion model should rely more on image conditioning and denoising refinement. We therefore ablate the co… 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

51 extracted references · 7 linked inside Pith

  1. [1]

    In: ECCV

    Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: ECCV. pp. 611–625 (2012)

    2012

  2. [2]

    Capito, L., Ozguner, U., Redmill, K.: Optical flow based visual potential field for autonomous driving. In: IV. pp. 885–891 (2020)

    2020

  3. [3]

    arXiv preprint arXiv:2606.26763 (2026)

    Chen, H., Zhang, X., Sun, X., Mingqing, X.: Calibrated harmonic overlaid implicit neural representations for multi-dimensional data. arXiv preprint arXiv:2606.26763 (2026)

    Pith/arXiv arXiv 2026

  4. [4]

    In: CVPR (2023)

    Dong, Q., Cao, C., Fu, Y.: Rethinking optical flow from geometric matching consistent perspective. In: CVPR (2023)

    2023

  5. [5]

    In: CVPR

    Dong, Q., Fu, Y.: Memflow: Optical flow estimation and prediction with memory. In: CVPR. pp. 19068–19078 (2024)

    2024

  6. [6]

    arXiv preprint arXiv:2312.01746 (2023)

    Dong, Q., Zhao, B., Fu, Y.: Open-ddvm: A reproduction and extension of diffusion model for optical flow estimation. arXiv preprint arXiv:2312.01746 (2023)

    Pith/arXiv arXiv 2023

  7. [7]

    In: ICCV

    Dosovitskiy, A., Fischer, P., Ilg, E., Hausser, P., Hazirbas, C., Golkov, V., Van Der Smagt, P., Cremers, D., Brox, T.: Flownet: Learning optical flow with convolu- tional networks. In: ICCV. pp. 2758–2766 (2015)

    2015

  8. [8]

    In: CVPR

    Fan, L., Huang, W., Gan, C., Ermon, S., Gong, B., Huang, J.: End-to-end learning of motion representation for video understanding. In: CVPR. pp. 6016–6025 (2018)

    2018

  9. [9]

    In: CVPR

    Gao, S., Liu, X., Zeng, B., Xu, S., Li, Y., Luo, X., Liu, J., Zhen, X., Zhang, B.: Implicit diffusion models for continuous super-resolution. In: CVPR. pp. 10021– 10030 (2023)

    2023

  10. [10]

    In: CVPR

    Hai, Y., Wang, G., Su, T., Jiang, W., Hu, Y.: Hierarchical flow diffusion for efficient frame interpolation. In: CVPR. pp. 22943–22952 (2025)

    2025

  11. [11]

    In: NeurIPS

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. pp. 6840–6851 (2020)

    2020

  12. [12]

    In: ECCV

    Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A transformer architecture for optical flow. In: ECCV. pp. 668–685. Springer (2022)

    2022

  13. [13]

    In: CVPR

    Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: Flownet 2.0: Evolution of optical flow estimation with deep networks. In: CVPR. pp. 2462–2470 (2017)

    2017

  14. [14]

    In: ICCV

    Ji, Y., Chen, Z., Xie, E., Hong, L., Liu, X., Liu, Z., Lu, T., Li, Z., Luo, P.: Ddp: Diffusion model for dense visual prediction. In: ICCV. pp. 21741–21752 (2023)

    2023

  15. [15]

    In: ICCV

    Jiang, S., Campbell, D., Lu, Y., Li, H., Hartley, R.: Learning to estimate hidden motions with global motion aggregation. In: ICCV. pp. 9772–9781 (2021)

    2021

  16. [16]

    In: ICCV

    Liu, R., Sun, S., Tang, H., Gao, W., Li, G.: Flow4agent: Long-form video under- standing via motion prior from optical flow. In: ICCV. pp. 23817–23827 (2025)

    2025

  17. [17]

    In: CVPR

    Luo, A., Li, X., Yang, F., Liu, J., Fan, H., Liu, S.: Flowdiffuser: Advancing optical flow estimation with diffusion models. In: CVPR. pp. 19167–19176 (2024)

    2024

  18. [18]

    In: ICCV

    Luo, A., Yang, F., Li, X., Nie, L., Lin, C., Fan, H., Liu, S.: Gaflow: Incorporating gaussian attention into optical flow. In: ICCV. pp. 9642–9651 (2023)

    2023

  19. [19]

    RA-L7, 5639–5646 (2022)

    Mahjourian, R., Kim, J., Chai, Y., Tan, M., Sapp, B., Anguelov, D.: Occupancy flow fields for motion forecasting in autonomous driving. RA-L7, 5639–5646 (2022)

    2022

  20. [20]

    In: CVPR

    Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: CVPR. pp. 4040–4048 (2016)

    2016

  21. [21]

    In: CVPR (2023) FlowPainter 17

    Mehl, L., Schmalfuss, J., Jahedi, A., Nalivayko, Y., Bruhn, A.: Spring: A high- resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In: CVPR (2023) FlowPainter 17

    2023

  22. [22]

    In: NeurIPS

    Meng, Y., Jin, X., Lei, L., Guo, C.L., Li, C.: Ultraled: Learning to see everything in ultra-high dynamic range scenes. In: NeurIPS. pp. 41466–41495 (2025)

    2025

  23. [23]

    In: CVPR (2015)

    Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: CVPR (2015)

    2015

  24. [24]

    In: CVPR

    Morimitsu, H., Zhu, X., Cesar, R.M., Ji, X., Yin, X.C.: Dpflow: Adaptive optical flow estimation with a dual-pyramid framework. In: CVPR. pp. 17810–17820 (2025)

    2025

  25. [25]

    In: ICCV

    Pepe, A., Dos Santos Mendonca, P., Lasenby, J.: Geometric inductive priors in diffusion-based optical flow estimation. In: ICCV. pp. 655–665 (2025)

    2025

  26. [26]

    In: CVPR

    Piergiovanni, A., Ryoo, M.S.: Representation flow for action recognition. In: CVPR. pp. 9945–9953 (2019)

    2019

  27. [27]

    In: ICCV

    Poggi, M., Tosi, F.: Flowseek: optical flow made easier with depth foundation models and motion bases. In: ICCV. pp. 5667–5679 (2025)

    2025

  28. [28]

    arXiv preprint arXiv:2606.27905 (2026)

    Qu, L., Liu, Y., Zhou, S., Liang, J., Zeng, H., Zhang, L., Yang, J.: There and back again: A flexible-frame transformer for multi-exposure fusion. arXiv preprint arXiv:2606.27905 (2026)

    Pith/arXiv arXiv 2026

  29. [29]

    In: ISVC

    Rakêt, L.L., Roholm, L., Bruhn, A., Weickert, J.: Motion compensated frame interpolation with a symmetric optical flow constraint. In: ISVC. pp. 447–457 (2012)

    2012

  30. [30]

    In: CVPR

    Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: CVPR. pp. 4161–4170 (2017)

    2017

  31. [31]

    In: NeurIPS

    Saxena, S., Herrmann, C., Hur, J., Kar, A., Norouzi, M., Sun, D., Fleet, D.J.: The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. In: NeurIPS. pp. 39443–39469 (2023)

    2023

  32. [32]

    arXiv preprint arXiv:2301.04422 (2023)

    Shen, S., Kerofsky, L., Yogamani, S.: Optical flow for autonomous driving: Applica- tions, challenges and improvements. arXiv preprint arXiv:2301.04422 (2023)

    Pith/arXiv arXiv 2023

  33. [33]

    In: CVPR

    Shi, X., Huang, Z., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Flowformer++: Masked cost volume autoencoding for pretraining optical flow estimation. In: CVPR. pp. 1599–1610 (2023)

    2023

  34. [34]

    arXiv preprint arXiv:2010.02502 (2020)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    Pith/arXiv arXiv 2010

  35. [35]

    In: CVPR

    Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: CVPR. pp. 8934–8943 (2018)

    2018

  36. [36]

    In: CVPR

    Sun, S., Kuang, Z., Sheng, L., Ouyang, W., Zhang, W.: Optical flow guided feature: A fast and robust motion representation for video action recognition. In: CVPR. pp. 1390–1399 (2018)

    2018

  37. [37]

    In: ECCV

    Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402–419 (2020)

    2020

  38. [38]

    In: CVPR

    Teed, Z., Deng, J.: Raft-3d: Scene flow using rigid-motion embeddings. In: CVPR. pp. 8375–8384 (2021)

    2021

  39. [39]

    Computational Visual Media11(1), 1–28 (2025)

    Wang, C., Peng, H.Y., Liu, Y.T., Gu, J., Hu, S.M.: Diffusion models for 3d genera- tion: A survey. Computational Visual Media11(1), 1–28 (2025)

    2025

  40. [40]

    IJCV132(12), 5929–5949 (2024)

    Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. IJCV132(12), 5929–5949 (2024)

    2024

  41. [41]

    In: ECCV

    Wang, Y., Lipson, L., Deng, J.: Sea-raft: Simple, efficient, accurate raft for optical flow. In: ECCV. pp. 36–54 (2024)

    2024

  42. [42]

    In: ECCV

    Wu, S., You, K., He, W., Yang, C., Tian, Y., Wang, Y., Zhang, Z., Liao, J.: Video interpolation by event-driven anisotropic adjustment of optical flow. In: ECCV. pp. 267–283 (2022)

    2022

  43. [43]

    In: ECCV

    Wulff, J., Butler, D.J., Stanley, G.B., Black, M.J.: Lessons and insights from creating a synthetic optical flow benchmark. In: ECCV. pp. 168–177 (2012)

    2012

  44. [44]

    In: CVPR

    Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: CVPR. pp. 8121–8130 (2022) 18 Y. Meng et al

    2022

  45. [45]

    IEEE TPAMI45(11), 13941–13958 (2023)

    Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Yu, F., Tao, D., Geiger, A.: Unifying flow, stereo and depth estimation. IEEE TPAMI45(11), 13941–13958 (2023)

    2023

  46. [46]

    arXiv preprint arXiv:2603.25739 (2026)

    Zhang, D., Wang, F., Pollefeys, M., Xu, H.: Megaflow: Zero-shot large displacement optical flow. arXiv preprint arXiv:2603.25739 (2026)

    Pith/arXiv arXiv 2026

  47. [47]

    In: ICCV

    Zhang, F., Woodford, O.J., Prisacariu, V.A., Torr, P.H.: Separable flow: Learning motion cost volumes for optical flow estimation. In: ICCV. pp. 10807–10817 (2021)

    2021

  48. [48]

    TOG43(4), 1–20 (2024)

    Zhang, L., Wang, Z., Zhang, Q., Qiu, Q., Pang, A., Jiang, H., Yang, W., Xu, L., Yu, J.: Clay: A controllable large-scale generative model for creating high-quality 3d assets. TOG43(4), 1–20 (2024)

    2024

  49. [49]

    In: ICCV

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836–3847 (2023)

    2023

  50. [50]

    arXiv preprint arXiv:2606.26769 (2026)

    Zhao, Q., Li, Y., Sun, Q., Yan, Z.: Resilphase: Plug-and-play phase mapping and noise-resilient macro-trajectory extrapolation for diffusion acceleration. arXiv preprint arXiv:2606.26769 (2026)

    Pith/arXiv arXiv 2026

  51. [51]

    In: CVPR

    Zhao, S., Sheng, Y., Dong, Y., Chang, E.I., Xu, Y., et al.: Maskflownet: Asymmetric feature matching with learnable occlusion mask. In: CVPR. pp. 6278–6287 (2020)

    2020

This paper was first reviewed by grok-4.5 on July 14, 2026.