Pith. sign in

REVIEW 2 major objections 7 minor 43 references

Stereo masks from one eye

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 · glm-5.2

2026-07-07 15:19 UTC pith:WQBQ6F6I

load-bearing objection Clean geometric theorem enables self-supervised stereo inpainting from monocular video; evaluation has a circularity concern that the paper partially addresses. the 2 major comments →

arxiv 2607.05354 v1 pith:WQBQ6F6I submitted 2026-07-06 cs.CV

Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation

Jingyi Lu , Kai Han This is my paper
classification cs.CV
keywords stereo inpaintingdisocclusionDIBRself-supervised learningcycle consistencymonocular depth estimationstereoscopic video generation
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.

The paper proves that the disocclusion mask needed to synthesize a right-eye view from a left-eye view equals the set of pixels lost when warping back from right to left. This means any single image can serve as its own training data for stereo inpainting: estimate its depth, analytically compute the mask of pixels that would be lost in a reverse warp, and use the original image as ground truth. The result is a self-supervised training pipeline that needs no stereo pairs, no synthetic data, and no expensive cycle computation.

Core claim

The Geometric Reciprocity Theorem (GRT): under nearest-neighbor DIBR warping, the disocclusion mask for left-to-right synthesis equals the lost-pixel mask from right-to-left warping, computable from monocular depth alone. This eliminates the need for paired stereo data, synthetic rendering, or multi-step cycle-consistency optimization in training stereo inpainting networks.

What carries the argument

A three-step simplification of the right-left-right cycle: (1) inpainted left-view content does not affect the backward disocclusion mask, so inpainting can be skipped; (2) disparity transfers through forward warping, so left-view depth estimation can be skipped; (3) pixels completing the forward-backward round trip return to their exact original positions, so warping itself can be skipped. The mask reduces to two analytical conditions on the target view: boundary violation and depth occlusion.

Load-bearing premise

The proof assumes nearest-neighbor warping with integer pixel coordinates, while real DIBR pipelines often use bilinear interpolation; the paper reports 99.5% mask agreement between the two but the discrepancy concentrates at depth discontinuities where disocclusion quality is most critical.

What would settle it

If one constructs a counterexample where a pixel completes the forward-backward round trip under nearest-neighbor warping but is nonetheless disoccluded in the left-to-right direction, the theorem would fail. More practically, if GRT-derived masks from monocular images produce systematically worse stereo inpainting than masks from true stereo pairs at test time, the claimed train-test equivalence would be empirically refuted.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Any monocular image or video becomes a training sample for stereo inpainting, removing the data scarcity bottleneck for stereoscopic content generation.
  • The approach generalizes across inpainting architectures, as shown by improvements on LaMa, Stable Diffusion variants, ProPainter, and StereoCrafter.
  • The framework is depth-estimator-agnostic: swapping depth models degrades performance gradually rather than catastrophically.
  • The released datasets (ImageNet-GRT, Kinetics-GRT, DAVIS-GRT) provide standardized benchmarks for evaluating stereo inpainting in isolation from the full DIBR pipeline.

Where Pith is reading between the lines

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

  • The GRT identity is a property of rectified stereo geometry and nearest-neighbor warping; analogous reciprocity relations may exist for other forward-backward warping pairs in structured geometric domains beyond stereo.
  • If monocular depth estimation continues to improve, the quality of GRT-derived masks will improve for free, narrowing the gap between self-supervised and fully supervised stereo training without changes to the framework.
  • The 0.5% mask discrepancy between bilinear and nearest-neighbor warping occurs at depth discontinuities, which are exactly where inpainting quality matters most; a hybrid approach using bilinear masks at boundaries could potentially recover that gap.

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

2 major / 7 minor

Summary. This paper introduces the Geometric Reciprocity Theorem (GRT), which shows that under nearest-neighbor DIBR warping, the disocclusion mask needed to synthesize a target view equals the mask of pixels lost when warping from target back to source. This enables self-supervised training of stereo inpainting networks from monocular images alone, without requiring paired stereo data or synthetic rendering. The three-step proof (Section 3.4.1) is clean: inpainted regions lack physical counterparts, disparity co-transfers with color during forward warping, and round-trip pixels return to original positions (Eq. 17). The authors construct ImageNet-GRT and Kinetics-GRT training datasets, evaluate on DAVIS-GRT and Inria 3DMovie, and demonstrate improvements over both training-free and supervised baselines.

Significance. The GRT derivation is genuinely parameter-free: the equivalence M^{L→R}_dis = M^{R→L}_lost (Eq. 12) follows from the warping equations (Eqs. 14–17) without fitted constants. The three-step proof is geometrically verifiable and each step is load-bearing. The practical payoff is substantial: it eliminates the need for stereo pairs or synthetic data in training stereo inpainting networks, replacing them with monocular images plus depth estimation. The broad applicability is demonstrated by fine-tuning multiple architectures (LaMa, SD variants, ProPainter, StereoCrafter) with consistent gains (Table 4). The pseudocode in Figure A1 is concrete and reproducible. The framework is depth-estimator-agnostic (Table A3) and scales with data (Table A5). The Inria 3DMovie evaluation (Table 3) provides an independent check using real stereo footage with viewing-comfort metrics, complementing the DAVIS-GRT evaluation that isolates the inpainting sub-task.

major comments (2)
  1. §A2, DAVIS-GRT description: The primary quantitative evaluation (Table 2) constructs both training data and evaluation masks using the same GRT procedure with the same depth estimator (Depth Anything V2). This creates a self-referential loop: if the depth estimator has systematic biases at object boundaries (where disocclusions matter most), those biases are baked into both training targets and evaluation ground truth, potentially inflating PSNR/SSIM. The paper does not test whether GRT-derived evaluation masks match true stereo disocclusions from real stereo pairs. The Inria 3DMovie evaluation (Table 3) partially addresses this by using SIoU and MEt3R on real stereo footage, but these are viewing-comfort metrics, not pixel-level inpainting quality metrics. Adding even a small-scale comparison against masks derived from real stereo pairs (with known stereo matching limitations) wouldstre
  2. §3.4.2: The claim of 'equivalence to full supervision' is precisely stated for the nearest-neighbor DIBR formulation, and the proof is correct under that formulation. However, the claim is presented in the abstract and introduction without the nearest-neighbor qualifier, which could mislead readers about the scope. The 0.5% mask disagreement between bilinear and nearest-neighbor warping (Table A2) occurs at depth discontinuities where disocclusion quality matters most. The paper should explicitly state the nearest-neighbor limitation in the abstract and main claims, and note that the 'equivalence' is formulation-specific rather than universal. This is a presentation issue that affects the central claim's scope.
minor comments (7)
  1. §3.4.1, Eq. 17: The round-trip argument x'_R = x_R is clean under nearest-neighbor warping but the equation should note that this holds only for pixels that complete the round-trip (i.e., non-lost pixels), not for all pixels. The current notation could be read as applying universally.
  2. Table 2: The inference time for Ours-Image (0.05s) is reported at 512×512, but training resolution is 256×256 (§A4.1). Clarify whether inference time is measured at training or evaluation resolution, and whether the model is evaluated at a different resolution than trained.
  3. §A2: The disparity scaling parameter α differs between training (0.1) and evaluation (0.06). This means the disocclusion mask distributions differ between training and evaluation. While the paper reports strong results, the choice of different α values should be motivated — why not use the same value? Does this affect the train-test consistency claim?
  4. Figure 2: The three-step diagram is informative but panel (ii) could clarify that the transferred disparity d̃_L is only valid at non-disoccluded pixels, which is the key insight enabling elimination of disparity re-estimation.
  5. §A9: The bilinear warping extension is a thoughtful addition. The 'subtle train-test consistency trade-off' is acknowledged but the practical impact is understated. The 0.5% disagreement (Table A2) at depth discontinuities could disproportionately affect perceptual quality. Consider adding a qualitative comparison of masks under both formulations at object boundaries.
  6. Table A2: The 'Cycle consistency, independently estimated' row shows 99.1% agreement, which is high but means ~1% of pixels differ. It would help to report what fraction of the differing pixels are at depth discontinuities, to contextualize whether the disagreement is concentrated where it matters.
  7. References: The paper cites several 2025–2026 works (e.g., Shvetsova et al. 2026, Lin et al. 2025, Yu et al. 2025b). Ensure these are correctly attributed and that arXiv preprints are marked as such, as some may not be peer-reviewed at the time of publication.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and positive assessment. We address both major comments below.

read point-by-point responses

  1. Referee: §A2, DAVIS-GRT description: The primary quantitative evaluation (Table 2) constructs both training data and evaluation masks using the same GRT procedure with the same depth estimator (Depth Anything V2). This creates a self-referential loop: if the depth estimator has systematic biases at object boundaries (where disocclusions matter most), those biases are baked into both training targets and evaluation ground truth, potentially inflating PSNR/SSIM. The paper does not test whether GRT-derived evaluation masks match true stereo disocclusions from real stereo pairs. The Inria 3DMovie evaluation (Table 3) partially addresses this by using SIoU and MEt3R on real stereo footage, but these are viewing-comfort metrics, not pixel-level inpainting quality metrics. Adding even a small-scale comparison against masks derived from real stereo pairs (with known stereo matching limitations) wouldstre

    Authors: The referee raises a valid concern about the self-referential structure of the DAVIS-GRT evaluation. We agree that using the same depth estimator (Depth Anything V2) for both training data construction and evaluation mask generation could, in principle, inflate PSNR/SSIM if the depth estimator has systematic boundary biases that align across the two. We will address this in two ways during revision. First, we will add an explicit discussion of this limitation in Section A2, acknowledging the self-referential risk and noting that the Inria 3DMovie evaluation (Table 3) serves as an independent check on real stereo footage, albeit with viewing-comfort rather than pixel-level inpainting metrics. Second, we will add a small-scale comparison on a subset of Inria 3DMovie frames where we compute disocclusion masks from real stereo pairs (using stereo matching) alongside GRT-derived masks, reporting their agreement (IoU). This will quantify the gap between GRT masks and stereo-matching-derived masks, complementing the existing Table A2 result showing 99.5% agreement between nearest-neighbor and bilinear warping. We note that stereo-matching-derived masks themselves contain errors (as documented in Section A8 and Figure A3), so this comparison reflects two imperfect but independent sources rather than a clean ground truth. Nevertheless, we agree the comparison is informative and will include it. revision: yes

  2. Referee: §3.4.2: The claim of 'equivalence to full supervision' is precisely stated for the nearest-neighbor DIBR formulation, and the proof is correct under that formulation. However, the claim is presented in the abstract and introduction without the nearest-neighbor qualifier, which could mislead readers about the scope. The 0.5% mask disagreement between bilinear and nearest-neighbor warping (Table A2) occurs at depth discontinuities where disocclusion quality matters most. The paper should explicitly state the nearest-neighbor limitation in the abstract and main claims, and note that the 'equivalence' is formulation-specific rather than universal. This is a presentation issue that affects the central claim's scope.

    Authors: We agree with the referee that the abstract and introduction should explicitly state the nearest-neighbor qualifier when claiming equivalence to full supervision. We note that the abstract already contains the phrase 'under the nearest-neighbor DIBR formulation' when first introducing GRT, and Section 3.4.2 is titled 'Equivalence to Full Supervision' with the opening sentence reading 'under the stated DIBR formulation.' However, the referee is correct that the phrase 'equivalence to full supervision' in the introduction (end of the paragraph following the GRT statement) does not repeat the qualifier, and a reader skimming could miss the scope limitation. We will revise the introduction to restate the nearest-neighbor qualifier when the equivalence claim is made, and we will add a sentence noting that the 0.5% disagreement under bilinear warping (Table A2) occurs at depth discontinuities, making the equivalence formulation-specific rather than universal. We will also add a forward reference to Section A9, which extends GRT to soft interpolation warping and discusses the train-test consistency trade-off. This is a presentation fix; the mathematical content and proof remain unchanged. revision: yes

Circularity Check

0 steps flagged

GRT derivation is a parameter-free geometric identity; no circular reduction found

full rationale

The paper's central result, M^{L→R}_dis = M^{R→L}_lost (Eq. 12), is proved in Section 3.4.1 through three independent geometric observations. The load-bearing step is Eq. 17: x'_R = x_L - d̃_L(x_L,y_L) = [x_R + d_R(x_R,y_R)] - [d_R(x_R,y_R)] = x_R. This is a straightforward algebraic identity: forward warping adds d_R, backward warping subtracts the same transferred d_R, so pixels completing the round-trip return to their original positions. The two masks are defined independently — M^{L→R}_dis by the warping operation W^{L→R} (Eq. 3) and M^{R→L}_lost by boundary violations and depth occlusion conditions (Eqs. 18–21) — and the theorem proves their equality. No parameter is fitted to make the theorem hold; α (0.1/0.06) is a data-construction hyperparameter, not a theorem parameter. No self-citation is load-bearing for the proof. The skeptic's concern about DAVIS-GRT using GRT-derived masks for both training and evaluation is a validity concern (depth estimator bias at boundaries), not a circularity in the derivation: the paper proves mask correctness via the theorem, then uses those masks to evaluate inpainting quality (PSNR/SSIM against original image content), not to re-prove mask correctness. The Inria 3DMovie evaluation uses an external stereo dataset with different metrics (SIoU, MEt3R), providing independent corroboration. Score 2 reflects the minor self-referential evaluation structure that does not undermine the theorem's independent mathematical content.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities, particles, forces, or postulated objects. The Geometric Reciprocity Theorem is a mathematical result, not an invented entity. The free parameters (α, depth model) are standard hyperparameters for data construction, not ad hoc constants introduced to make the derivation work. The axioms are all domain assumptions specific to the DIBR framework, clearly stated in the paper, and none are invented for the paper.

free parameters (2)
  • α (disparity scaling) = 0.1 (training), 0.06 (evaluation)
    Controls the range of disparity values after rescaling relative depth to [0, αW]. Not fitted to optimize the theorem but chosen to match typical stereo baseline ranges. Affects mask size and content.
  • Depth model choice = Depth Anything V2-Large
    The specific depth estimator determines the disparity maps from which all masks are computed. The paper shows robustness to alternative estimators (Table A3) but the choice affects results.
axioms (4)
  • domain assumption Nearest-neighbor warping: each pixel maps to a discrete integer coordinate via rounding.
    The GRT proof in Section 3.4.1 requires this for the exact round-trip property (Eq. 17). Bilinear warping introduces an approximation gap (Section A9).
  • domain assumption Rectified stereo geometry: vertical coordinates are preserved during warping (y' = y).
    Stated in Eq. 2 and assumed throughout. Non-rectified stereo would require epipolar transfer and break the 1D warping formulation.
  • domain assumption Depth occlusion follows a winner-take-all rule: the nearest pixel (max disparity) is retained at each target location.
    Used in Eq. 19 for the depth buffer. Standard in DIBR but assumes opaque, non-transparent surfaces. The paper acknowledges transparent/reflective surfaces as a limitation (Section A6).
  • domain assumption Monocular depth estimation provides sufficiently accurate disparity at object boundaries to produce correct disocclusion masks.
    The entire framework depends on depth quality. The paper shows depth-estimator robustness (Table A3) but does not bound the error from depth inaccuracy at discontinuities.

pith-pipeline@v1.1.0-glm · 19860 in / 3197 out tokens · 295366 ms · 2026-07-07T15:19:39.433105+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation." pith.science (2026). https://pith.science/paper/WQBQ6F6I

@misc{pith2026260705354,
  author       = {Pith},
  title        = {Pith review of: Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WQBQ6F6I}},
  note         = {Machine review of arXiv:2607.05354}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Monocular-to-stereo conversion synthesizes stereoscopic content from 2D videos for immersive 3D experiences. In modern Depth-Image-Based Rendering (DIBR) approaches, stereo inpainting of disocclusions is the critical bottleneck. Training-based methods achieve superior quality but rely on scarce stereo pairs or synthetic data with domain gaps. We address this through the first self-supervised framework learning from monocular videos via cycle consistency. Our key contribution is the Geometric Reciprocity Theorem (GRT): under the nearest-neighbor DIBR formulation, the disocclusion mask when synthesizing a target view equals the mask of pixels lost when warping back from target to source, enabling analytical computation of test-time disocclusion masks directly from monocular images. This yields train-test consistency for the stated warping formulation, supporting self-supervised learning from unlimited monocular videos and substantial improvements over training-free and supervised state-of-the-art methods. Project page: https://visual-ai.github.io/grt/

Figures

Figures reproduced from arXiv: 2607.05354 by Jingyi Lu, Kai Han.

Figure 1
Figure 1. Figure 1: Cycle consistency framework. Given right view IR, the complete cycle synthesizes left view through depth estimation (D), forward warping (WR→L), and inpainting (G), then reconstructs the right view via depth estimation on the synthesized left view, backward warping (WL→R), and inpainting. The reconstruction loss Lcycle = ∥Iˆrecon R − IR∥ provides self-supervision. videos predominantly come from copyrighted… view at source ↗
Figure 2
Figure 2. Figure 2: Progressive simplification of cycle consistency to Geometric Reciprocity. (i) Inpainted regions in IˆL do not affect Iˆrecon R , allowing us to skip left view inpainting (marked with ×) and directly use I˜L. (ii) Right-to-left warping transfers disparity from dR to I˜L, allowing us to skip left view disparity estimation and directly reuse d˜L. (iii) Transferred disparity ensures perfect round-trips for all… view at source ↗
Figure 3
Figure 3. Figure 3: Equivalence to Full Supervision. Top: inference synthesizes IR from IL via DIBR and obtains ML→R dis for inpainting. Bottom: GRT treats a monocular frame as IR and computes the same mask analytically from IR alone for training and DAVIS-GRT evaluation. 3.5. Model Architecture We adopt LaMa (Suvorov et al., 2022), which leverages Fast Fourier Convolutions, for image stereo inpainting and ProPainter (Zhou et… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison. Our method produces more natural textures and smoother boundaries [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] 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

43 extracted references · 43 canonical work pages · 2 internal anchors

  1. [1]

    Deep3d: Fully Automatic 2d-to-3d Video Conversion with Deep Convolutional Neural Networks , author =

    work page

  2. [2]

    Shvetsova, Nina and Bhat, Goutam and Truong, Prune and Kuehne, Hilde and Tombari, Federico , booktitle =

    work page

  3. [3]

    Zhao, Sijie and Hu, Wenbo and Cun, Xiaodong and Zhang, Yong and Li, Xiaoyu and Kong, Zhe and Gao, Xiangjun and Niu, Muyao and Shan, Ying , journal =

    work page

  4. [4]

    Shi, Jian and Wang, Qian and Li, Zhenyu and Idoughi, Ramzi and Wonka, Peter , journal =

    work page

  5. [5]

    Stereodiffusion: Training-Free Stereo Image Generation Using Latent Diffusion Models , author =

    work page

  6. [6]

    Huang, Xingchang and Singh, Ashish Kumar and Dubost, Florian and Vasconcelos, Cristina Nader and Khattar, Sakar and Shi, Liang and Theobalt, Christian and Oztireli, Cengiz and Singh, Gurprit , journal =

    work page

  7. [7]

    Dai, Peng and Tan, Feitong and Xu, Qiangeng and Futschik, David and Du, Ruofei and Fanello, Sean and Qi, Xiaojuan and Zhang, Yinda , journal =

    work page

  8. [8]

    Are We Ready for Autonomous Driving? The KITTI Vision Benchmark Suite , author =

    work page

  9. [9]

    Indoor Segmentation and Support Inference from RGBD Images , author =

    work page

  10. [10]

    Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer , author =

    work page

  11. [11]

    Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data , author =

    work page

  12. [12]

    Depth Pro: Sharp Monocular Metric Depth in Less Than a Second , author =

    work page

  13. [13]

    Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation , author =

    work page

  14. [14]

    Vision Transformers for Dense Prediction , author =

    work page

  15. [15]

    Kong, Lingdong and Xie, Shaoyuan and Hu, Hanjiang and Ng, Lai Xing and Cottereau, Benoit and Ooi, Wei Tsang , booktitle = NIPS, year =

    work page

  16. [16]

    Zheng, Yupeng and Zhong, Chengliang and Li, Pengfei and Gao, Huan-ang and Zheng, Yuhang and Jin, Bu and Wang, Ling and Zhao, Hao and Zhou, Guyue and Zhang, Qichao and others , booktitle = ICRA, year =

    work page

  17. [17]

    Depth Anything at Any Condition

    Depth Anything at Any Condition , author =. arXiv preprint arXiv:2507.01634 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks , author =

    work page

  19. [19]

    ICML , year =

    CyCADA: Cycle-Consistent Adversarial Domain Adaptation , author =. ICML , year =

    work page

  20. [20]

    Temporal Cycle-Consistency Learning , author =

    work page

  21. [21]

    Self-Supervised Video Object Segmentation by Motion Grouping , author =

    work page

  22. [22]

    Learning Dense Correspondence via 3D-Guided Cycle Consistency , author =

    work page

  23. [23]

    Unsupervised Image Super-Resolution Using Cycle-in-Cycle Generative Adversarial Networks , author =

    work page

  24. [24]

    Singh, Ankit , booktitle = NIPS, year =

    work page

  25. [25]

    Kubric: A scalable dataset generator , author =

    work page

  26. [26]

    Russakovsky, Olga and Deng, Jia and Su, Hao and Krause, Jonathan and Satheesh, Sanjeev and Ma, Sean and Huang, Zhiheng and Karpathy, Andrej and Khosla, Aditya and Bernstein, Michael and others , journal = IJCV, year =

    work page

  27. [27]

    Kay, Will and Carreira, Joao and Simonyan, Karen and Zhang, Brian and Hillier, Chloe and Vijayanarasimhan, Sudheendra and Viola, Fabio and Green, Tim and Back, Trevor and Natsev, Paul and others , journal =. The

    work page

  28. [28]

    A benchmark dataset and evaluation methodology for video object segmentation , author =

    work page

  29. [29]

    Youtube-vos: Sequence-to-sequence video object segmentation , author =

    work page

  30. [30]

    Resolution-Robust Large Mask Inpainting with

    Suvorov, Roman and Logacheva, Elizaveta and Mashikhin, Anton and Remizova, Anastasia and Ashukha, Arsenii and Silvestrov, Aleksei and Kong, Naejin and Goka, Harshith and Park, Kiwoong and Lempitsky, Victor , booktitle = WACV, year =. Resolution-Robust Large Mask Inpainting with

    work page

  31. [31]

    Image Quality Assessment: From Error Visibility to Structural Similarity , author =

    work page

  32. [32]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author =

    work page

  33. [33]

    Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle = ICLR, year =

    work page

  34. [34]

    High-Resolution Image Synthesis with Latent Diffusion Models , author =

    work page

  35. [35]

    TrajectoryCrafter: Redirecting Camera Trajectory for Monocular Videos via Diffusion Models , author =

    work page

  36. [36]

    Propainter: Improving propagation and transformer for video inpainting , author=

    work page

  37. [37]

    Min, Junhong and Kim, Jimin and Min, Cheol-Hui and Kim, Minwook and Jeon, Youngpil and Choi, Minyong , journal=

    work page

  38. [38]

    Depth Anything 3: Recovering the Visual Space from Any Views

    Depth Anything 3: Recovering the Visual Space from Any Views , author=. arXiv preprint arXiv:2511.10647 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    Video depth anything: Consistent depth estimation for super-long videos , author=

    work page

  40. [40]

    Mono2Stereo: A Benchmark and Empirical Study for Stereo Conversion , author=

    work page

  41. [41]

    Wang, Xianqi and Yang, Hao and Xu, Gangwei and Cheng, Junda and Lin, Min and Deng, Yong and Zang, Jinliang and Chen, Yurui and Yang, Xin , booktitle=ICCV, year=

    work page

  42. [42]

    Pose estimation and segmentation of people in 3D movies , author=

    work page

  43. [43]

    Met3r: Measuring multi-view consistency in generated images , author=

    work page