Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

CamCloneMaster: Enabling Reference-based Camera Control for Video Generation

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CamCloneMaster claims that a video diffusion transformer can clone camera motion directly from a reference video by concatenating latent tokens, matching or beating parameter-conditioned baselines without camera parameters or test-time…

desk verdict Useful reference-based camera cloning with a strong dataset, but the SOTA claim rests on unvalidated MegaSaM metrics and a contradictory protocol description. read the letter →

arxiv 2506.03140 v1 pith:Q67QZNR3 submitted 2025-06-03 cs.CV

Yawen Luo , Jianhong Bai , Xiaoyu Shi , Menghan Xia , Xintao Wang , Pengfei Wan , Di Zhang , Kun Gai
show 1 more author
Tianfan Xue
This is my paper
classification cs.CV
keywords cameracontrolreference-basedvideogenerationdiffusiontransformerimage-to-videovideo-to-videoregenerationlatenttokenconcatenationsyntheticpaireddatasetmotioncloning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Camera control in video generation usually starts from explicit pose sequences, which are hard to write and fragile to estimate. This paper claims a more direct route: give the model a reference clip, and it will replicate that camera trajectory in a newly generated video, with no camera parameters and no per-video fine-tuning. The method concatenates the reference video's VAE latents with the noisy video tokens along the frame dimension and fine-tunes only the 3D spatial-temporal attention layers of a latent diffusion transformer on a large synthetic paired dataset. The authors report that this outperforms parameter-conditioned baselines, even baselines given ground-truth poses, on camera accuracy, visual quality, and dynamic quality, in both image-to-video and video-to-video settings. If true, the result makes cinematic camera control as intuitive as supplying an example shot.

What carries the argument

The load-bearing mechanism is token concatenation in the frame dimension. The reference video (and, in V2V, the content video) is encoded by the same 3D VAE used for the target, patchified into tokens, and concatenated with the noisy target tokens before the transformer, so the 3D spatial-temporal attention layers can model the interaction between condition and noise tokens without any new control module. Only those attention layers are fine-tuned, which preserves the base model's generative ability while teaching it camera cloning. The supervision comes from a large synthetic dataset of paired videos in which synchronized cameras record the same scenes and characters from identical trajectories across different locations, giving the model explicit examples of 'same trajectory, different content.'

What would settle it

Render a held-out set of synthetic scenes with known camera trajectories, generate videos with CamCloneMaster and with parameter-based baselines using exact ground-truth poses, and compare each output's extracted trajectory against the simulator's known path; if CamCloneMaster's rotation and translation advantage shrinks or reverses under this direct measurement, the claimed ordering depends on the evaluation pose estimator.

Watch

Extended reading notes

Core claim

The paper's central claim is that camera motion can be learned as a transferable property of video latents. A diffusion transformer conditioned by frame-dimension token concatenation of a reference video's latent tokens learns to reproduce the reference camera trajectory in a new scene, and the same mechanism handles two tasks at once: animating a starting image with the reference motion, and re-shooting a supplied content video with that motion. The authors train only the 3D spatial-temporal attention layers, on paired synthetic videos that share camera trajectories across different locations and characters, and then evaluate on real-world image and video references. They report the lowest rotation error, translation error, and camera-motion consistency among compared methods, and a user study in which most participants prefer its camera accuracy and temporal consistency over parameter-based and training-free baselines.

Load-bearing premise

The load-bearing premise is that the pose estimator used to score camera accuracy gives unbiased enough trajectories to rank methods by true camera fidelity, even though the paper's own user study shows parameter-based methods are extremely sensitive to estimated-pose quality.

Editorial extensions

If this is right

  • A user can specify a shot's camera motion by supplying any reference clip, removing the need to estimate or design camera parameter sequences.
  • The same trained model serves both image-to-video and video-to-video generation, since the content reference is just another token sequence that can be replaced by zeros in the I2V case.
  • Camera fidelity no longer has to be capped by inference-time pose estimation: the paper's user study shows parameter-based methods lose most user preference when fed estimated poses instead of ground truth.
  • Because only the 3D spatial-temporal attention layers are tuned, camera cloning can be added to an existing latent diffusion transformer with a relatively small parameter footprint.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same frame-concatenation recipe may extend to other global video attributes, such as lighting style or object-motion style, provided paired synthetic data with identical attributes can be rendered.
  • The reported camera-accuracy ordering should be tested against independent trajectory measurements; since the metrics rely on a single pose estimator, a held-out synthetic set with known ground-truth trajectories would be a sharper test.
  • Training on only rendered scenes means real-world generalization is demonstrated qualitatively; a small-scale real paired video fine-tune could quantify how much synthetic-to-real transfer actually contributes.
  • The paper's stated limitation is that token concatenation raises computational demand, with sparse attention and latent drop left as future work, so the practical efficiency of the approach is not yet established.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper introduces CamCloneMaster, a training-based framework for reference-based camera control in video generation. It injects a camera-motion reference video (and optionally a content-reference video) into a latent diffusion transformer by concatenating VAE-encoded tokens along the frame dimension, fine-tuning only the 3D spatial-temporal attention layers. The authors also contribute a large synthetic Unreal Engine dataset of paired videos with shared camera trajectories, and report quantitative, qualitative, and user-study results for both image-to-video (I2V) and video-to-video (V2V) settings, claiming state-of-the-art camera accuracy and visual quality.

Significance. If the results hold, the contribution is practically valuable: it removes the need for explicit camera parameters or test-time fine-tuning, unifies I2V and V2V camera control with one simple architectural change, and the released dataset could support future research. The paper is clearly written, the token-concatenation design is simple, and the ablations in Tables 5 and 6 support the main architectural choices. The authors also provide a user study and report many standard metrics. However, the headline camera-accuracy numbers rest on MegaSaM-estimated trajectories without validation on ground-truth data, and all tables lack error bars, so the state-of-the-art claim needs stronger support.

major comments (5)
  1. [Sec. 5.1, Table 1] The RotErr, TransErr, and CamMC metrics are computed from camera trajectories extracted by MegaSaM on the generated clips, but the paper never validates MegaSaM against the ground-truth camera parameters that RealEstate10K provides for the evaluation references. Table 3 is direct evidence that MegaSaM-estimated poses are substantially worse than ground truth for downstream camera fidelity (e.g., CameraCtrl user preference drops from 87.67% to 13.33%). Since CamCloneMaster is trained to reproduce pixel-level apparent motion, MegaSaM's estimation error may be method-dependent and could systematically favor the proposed method. Please report MegaSaM's error on the RealEstate10K reference videos relative to their ground-truth poses, compute the camera-accuracy metrics on a subset where MegaSaM is calibrated, and provide confidence intervals or bootstrap estimates.
  2. [Tables 1, 2, 5, 6] All quantitative tables report single-run values without error bars, significance tests, or repeated sampling. The margins over the best baseline in Table 1 are modest (RotErr 1.49 vs. 1.62; TransErr 2.37 vs. 3.07; CamMC 3.50 vs. 4.22) and could change with sampling noise. Please report results over multiple seeds or sampling runs with means and standard deviations, and perform significance tests, especially for the camera-accuracy and FVD/FID metrics.
  3. [Sec. 5.2.1 and Fig. 4 caption] The evaluation protocol for parameter-based baselines is ambiguous. Section 5.2.1 states that CameraCtrl and CamI2V receive ground-truth camera poses for Table 1, but Figure 4's caption says 'Camera poses are estimated using MegaSam for parameter-based methods.' Please clarify which results use ground-truth versus estimated poses, and fix the apparent typo in the figure caption, which says 'Quantitative Results' where 'Qualitative Results' is evidently intended.
  4. [Sec. 4 and Sec. 5] The real-world generalization claim rests on training with only 40 rendered Unreal Engine scenes, with qualitative demonstrations on a handful of clips. While the RealEstate10K evaluation uses real reference videos, the paper does not analyze the sim-to-real gap or provide quantitative evidence on a larger real-world set. Please include a quantitative analysis of domain shift (e.g., performance across reference-video categories with different dynamic content, or an evaluation on a held-out real-world set) or temper the generalization claim accordingly.
  5. [Table 5] The injection-mechanism ablation compares token concatenation against a ControlNet-like architecture with duplicated DiT blocks, which has substantially more parameters and different training cost. The conclusion that 'token concatenation outperforms feature addition' may therefore be confounded by capacity or optimization budget. Please report parameter counts and training FLOPs for each variant, or control for them, so that the ablation isolates the injection mechanism.
minor comments (6)
  1. [Throughout] The name 'MegaSaM' is spelled inconsistently; for example, Table 3 and Figure 4 use 'MegaSam.' Please unify the spelling.
  2. [Fig. 1 caption] The description of the V2V content reference ('the downsized content reference video is positioned beside the prompt') is unclear; please clarify how the content reference video is fed to the model and how it relates to the token concatenation in Fig. 2.
  3. [Sec. 3.2, Eq. (5)] The notation 'Frame Concat' is not formally defined with tensor axes; please specify the exact dimensions being concatenated (e.g., f, h, w, c) and the order of x_t, x_cam, and x_cont in the sequence.
  4. [Sec. 4] The relationship between 391K videos, 39.1K locations, and 1,155K triple video sets is not derived; please explain the counting procedure for constructing the triple sets.
  5. [Sec. 5.1] The exact number of generated videos used for computing FVD and FID, the frame sampling protocol, and any confidence thresholds used in the VBench metrics are not specified; please provide these details for reproducibility.
  6. [Sec. 4] The paper states that the Camera Clone Dataset will be publicly released but provides no URL or release timeline; please add dataset availability details.

Circularity Check

0 steps flagged · score 0.0 of 10

No construction-level circularity: the camera-clone model is trained with a standard flow-matching loss and evaluated on external benchmarks, with no evaluation metric appearing in the training objective.

full rationale

The claimed derivation chain is self-contained in the relevant sense. CamCloneMaster's training signal is the flow-matching MSE in Eq. 2 over latent video tokens, not any of the headline camera-accuracy metrics (RotErr, TransErr, CamMC). The conditioning mechanism is token concatenation (Eq. 5) of VAE latents, and the model is fine-tuned on a synthetic paired dataset rendered in Unreal Engine. Evaluation is performed on external RealEstate10K and Koala-36M content with independently published baselines (CameraCtrl, CamI2V, MotionClone, DaS, ReCamMaster, TrajectoryCrafter), so the headline numbers are not fitted to the target quantities. The use of MegaSaM to extract camera trajectories for RotErr/TransErr/CamMC is an evaluation-validity concern, not a circularity: the metric is not part of the training loss, the reference trajectories are not defined in terms of the model's output, and the ordering is not forced by construction. Several cited works share authors with this paper (e.g., ReCamMaster, Cinemaster, Koala-36M), but they appear as baselines, data sources, or architectural reference points, not as the load-bearing justification for the central camera-cloning claim. No step in the derivation reduces by construction to its own input.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper is an empirical systems contribution, so the ledger items are hand-chosen hyperparameters and domain assumptions rather than fitted physical constants. No new entities are postulated. The most consequential entries are the domain assumptions: synthetic-to-real transfer and the use of an external pose estimator as ground truth for the headline camera-accuracy numbers.

free parameters (5)
  • I2V/V2V training mix = 50% / 50%
    Sec. 3.3 sets a balanced 50/50 split to give one model both abilities; no sensitivity analysis is reported, and the unified behavior depends on this balance.
  • Finetune scope = 3D spatial-temporal attention layers only
    Sec. 3.3 and ablation Table 6; freezing all other weights is a design choice on which the camera clone accuracy and visual quality results depend.
  • Training schedule = 12,000 steps, batch 64, lr 5e-5, 64 H800 GPUs
    Sec. 5.1 Implementation Details; hand-chosen and reported without sensitivity analysis.
  • Input sampling = 384x672 resolution, 77 frames
    Sec. 5.1; all metrics are measured at this scale, so the SOTA comparisons are tied to this resolution and length.
  • Dataset composition = 40 scenes, 66 characters, 10 trajectories per location group
    Sec. 4; the dataset's diversity is hand-configured, and the claimed generalization to real scenes is only verified qualitatively.
assumptions (5)
  • standard math Rectified-flow objective (Eq. 1-2) is the correct training target
    Sec. 3.1 adopts flow matching from Lipman et al. 2023 and Liu et al. 2022; used as-is.
  • domain assumption Camera motion information survives VAE encoding and token concatenation
    Sec. 3.2, Eq. 3-5; the entire method presumes the base 3D VAE latent of a reference video retains enough motion signal for the 3D attention layers to clone it.
  • domain assumption Synthetic UE5 renderings transfer to real-world camera motion
    Sec. 4; training uses only 40 rendered scenes with stylized characters, yet the model is evaluated on real RealEstate10K and internet clips.
  • domain assumption MegaSaM estimates are a valid ground truth for camera accuracy
    Sec. 5.1; RotErr, TransErr, and CamMC are computed from MegaSaM-estimated poses on generated videos, with no ground-truth verification on real data.
  • domain assumption Base model generative priors survive selective fine-tuning
    Sec. 3.3; freezing most layers assumes the internal I2V base model's quality is preserved, verified only through aggregate metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CamCloneMaster: Enabling Reference-based Camera Control for Video Generation." pith.science (2026). https://pith.science/paper/Q67QZNR3

@misc{pith2026250603140,
  author       = {Pith},
  title        = {Pith review of: CamCloneMaster: Enabling Reference-based Camera Control for Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q67QZNR3}},
  note         = {Machine review of arXiv:2506.03140}
}
read the original abstract

Camera control is crucial for generating expressive and cinematic videos. Existing methods rely on explicit sequences of camera parameters as control conditions, which can be cumbersome for users to construct, particularly for intricate camera movements. To provide a more intuitive camera control method, we propose CamCloneMaster, a framework that enables users to replicate camera movements from reference videos without requiring camera parameters or test-time fine-tuning. CamCloneMaster seamlessly supports reference-based camera control for both Image-to-Video and Video-to-Video tasks within a unified framework. Furthermore, we present the Camera Clone Dataset, a large-scale synthetic dataset designed for camera clone learning, encompassing diverse scenes, subjects, and camera movements. Extensive experiments and user studies demonstrate that CamCloneMaster outperforms existing methods in terms of both camera controllability and visual quality.

Figures

Figures reproduced from arXiv: 2506.03140 by the authors.

Figure 1
Figure 1. Camera Control results of CamCloneMaster. CamCloneMaster is capable of cloning camera motion from reference videos [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed CamCloneMaster. Given a camera motion reference video and an optional content reference video as [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Dataset Construction Illustration. We collect several [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Quantitative Results for Camera-Controlled Image-to-Video Generation. Camera poses are estimated using MegaSam for [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Quantitative Results for Camera-Controlled V2V Re-Generation. Camera poses are estimated using MegaSam for parameter [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CustomX: Unified Character, Action, and Scene Customization in Video World Models

    cs.CV 2025-12 conditional novelty 6.0 of 10

    AniX generates controllable videos of a user-supplied character performing typed actions inside a user-supplied 3D scene by fine-tuning a pre-trained video generator on small locomotion datasets.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [2]

    Lindell, and Sergey Tulyakov

    Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Ali- aksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B. Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion trans- formers, 2025. 2

    work page 2025

  2. [3]

    Lindell, and Sergey Tulyakov

    Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siaro- hin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, David B. Lindell, and Sergey Tulyakov. Vd3d: Taming large video diffusion transformers for 3d camera control, 2025

    work page 2025

  3. [4]

    Recammaster: Camera-controlled gen- erative rendering from a single video, 2025

    Jianhong Bai, Menghan Xia, Xiao Fu, Xintao Wang, Lianrui Mu, Jinwen Cao, Zuozhu Liu, Haoji Hu, Xiang Bai, Pengfei Wan, and Di Zhang. Recammaster: Camera-controlled gen- erative rendering from a single video, 2025. 2, 3, 5, 6, 7, 8

    work page 2025

  4. [5]

    Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking, 2025

    Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yijin Li, Fu- Yun Wang, and Hongsheng Li. Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking, 2025. 3

    work page 2025

  5. [6]

    Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 3

    2023

  6. [7]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luh- man, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators

  7. [8]

    Leap-vo: Long-term effective any point tracking for visual odometry, 2024

    Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry, 2024. 3

    work page 2024

  8. [9]

    Unreal engine 5

    Epic Games. Unreal engine 5. https://www.unrealengine.com/zh-CN/unreal-engine-5. 2, 4

    work page

Show all 49 references
  1. [10]

    Diffusion as shader: 3d- aware video diffusion for versatile video generation control,

    Zekai Gu, Rui Yan, Jiahao Lu, Peng Li, Zhiyang Dou, Chenyang Si, Zhen Dong, Qifeng Liu, Cheng Lin, Ziwei Liu, Wenping Wang, and Yuan Liu. Diffusion as shader: 3d- aware video diffusion for versatile video generation control,

  2. [11]

    Cameractrl: Enabling camera control for text-to-video generation, 2025

    Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation, 2025. 2, 3, 5, 6, 7, 8

    2025

  3. [12]

    Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models, 2025

    Hao He, Ceyuan Yang, Shanchuan Lin, Yinghao Xu, Meng Wei, Liangke Gui, Qi Zhao, Gordon Wetzstein, Lu Jiang, and Hongsheng Li. Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models, 2025. 3

    2025

  4. [13]

    Latent video diffusion models for high-fidelity long video generation, 2023

    Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 3

    2023

  5. [14]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium, 2018

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium, 2018. 5

    2018

  6. [15]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffu- sion models, 2022. 3

    2022

  7. [16]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers, 2022

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers, 2022. 3

    2022

  8. [17]

    Generative camera dolly: Ex- treme monocular dynamic novel view synthesis, 2024

    Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sar- gent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl V ondrick. Generative camera dolly: Ex- treme monocular dynamic novel view synthesis, 2024. 3

    2024

  9. [18]

    Training-free camera control for video generation, 2025

    Chen Hou and Zhibo Chen. Training-free camera control for video generation, 2025. 2, 3

    2025

  10. [19]

    Mo- tionmaster: Training-free camera motion transfer for video generation, 2024

    Teng Hu, Jiangning Zhang, Ran Yi, Yating Wang, Hongrui Huang, Jieyu Weng, Yabiao Wang, and Lizhuang Ma. Mo- tionmaster: Training-free camera motion transfer for video generation, 2024. 2, 3

    2024

  11. [20]

    Vbench: Com- prehensive benchmark suite for video generative models,

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench: Com- prehensive benchmark suite for video generative models,

  12. [21]

    Auto-encoding varia- tional bayes, 2022

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes, 2022. 3

    2022

  13. [22]

    Realcam-i2v: Real- world image-to-video generation with interactive complex camera control, 2025

    Teng Li, Guangcong Zheng, Rui Jiang, Shuigenzhan, Tao Wu, Yehao Lu, Yining Lin, and Xi Li. Realcam-i2v: Real- world image-to-video generation with interactive complex camera control, 2025. 2, 3

    2025

  14. [23]

    Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos, 2024

    Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holyn- ski, and Noah Snavely. Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos, 2024. 2, 3, 6, 7

    2024

  15. [24]

    Motionclone: Training-free motion cloning for controllable video generation, 2024

    Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation, 2024. 2, 3, 5, 6, 8

    2024

  16. [25]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maxi- milian Nickel, and Matt Le. Flow matching for generative modeling, 2023. 3

    2023

  17. [26]

    Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. 3

    2022

  18. [27]

    Latte: La- tent diffusion transformer for video generation, 2025

    Xin Ma, Yaohui Wang, Xinyuan Chen, Gengyun Jia, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: La- tent diffusion transformer for video generation, 2025. 3

    2025

  19. [28]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 3

    2023

  20. [29]

    Gen3c: 3d-informed world-consistent video generation with precise camera con- trol, 2025

    Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas M ¨uller, Alexan- der Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world-consistent video generation with precise camera con- trol, 2025. 8 9

    2025

  21. [30]

    High-resolution image syn- thesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models, 2022. 3

    2022

  22. [31]

    Gim: Learning generalizable image matcher from internet videos,

    Xuelun Shen, Zhipeng Cai, Wei Yin, Matthias M ¨uller, Zijun Li, Kaixuan Wang, Xiaozhi Chen, and Cheng Wang. Gim: Learning generalizable image matcher from internet videos,

  23. [32]

    Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling, 2024

    Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Che- ung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling, 2024. 3

    2024

  24. [33]

    Freeman, Joshua B

    Vincent Sitzmann, Semon Rezchikov, William T. Freeman, Joshua B. Tenenbaum, and Fredo Durand. Light field net- works: Neural scene representations with single-evaluation rendering, 2022. 3

    2022

  25. [34]

    Lightmotion: A light and tuning-free method for simulating camera motion in video generation, 2025

    Quanjian Song, Zhihang Lin, Zhanpeng Zeng, Ziyue Zhang, Liujuan Cao, and Rongrong Ji. Lightmotion: A light and tuning-free method for simulating camera motion in video generation, 2025. 3

    2025

  26. [35]

    To- wards accurate generative models of video: A new metric & challenges, 2019

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. To- wards accurate generative models of video: A new metric & challenges, 2019. 5

    2019

  27. [36]

    Cinemaster: A 3d-aware and controllable framework for cinematic text-to-video generation, 2025

    Qinghe Wang, Yawen Luo, Xiaoyu Shi, Xu Jia, Huchuan Lu, Tianfan Xue, Xintao Wang, Pengfei Wan, Di Zhang, and Kun Gai. Cinemaster: A 3d-aware and controllable framework for cinematic text-to-video generation, 2025. 3, 8

    2025

  28. [37]

    Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content, 2025

    Qiuheng Wang, Yukai Shi, Jiarong Ou, Rui Chen, Ke Lin, Jiahao Wang, Boyuan Jiang, Haotian Yang, Mingwu Zheng, Xin Tao, Fei Yang, Pengfei Wan, and Di Zhang. Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content, 2025. 5

    2025

  29. [38]

    Dust3r: Geometric 3d vi- sion made easy, 2024

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy, 2024. 2

    2024

  30. [39]

    Motionctrl: A unified and flexible motion controller for video generation,

    Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation,

  31. [40]

    Wan: Open and advanced large-scale video generative models, 2025

    WanTeam, :, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jin- gren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang,...

    2025

  32. [41]

    Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency, 2025

    Yiming Xie, Chun-Han Yao, Vikram V oleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency, 2025. 6

    2025

  33. [42]

    Dynamicrafter: Animating open-domain im- ages with video diffusion priors, 2023

    Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamicrafter: Animating open-domain im- ages with video diffusion priors, 2023. 3

    2023

  34. [43]

    Camco: Camera- controllable 3d-consistent image-to-video generation, 2024

    Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera- controllable 3d-consistent image-to-video generation, 2024. 2, 3

    2024

  35. [44]

    Tra- jectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models, 2025

    Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Tra- jectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models, 2025. 2, 3, 5, 6, 7, 8

    2025

  36. [45]

    Jacobs, Yael Pritch, Inbar Mosseri, Mike Zheng Shou, Neal Wadhwa, and Nataniel Ruiz

    David Junhao Zhang, Roni Paiss, Shiran Zada, Nikhil Karnad, David E. Jacobs, Yael Pritch, Inbar Mosseri, Mike Zheng Shou, Neal Wadhwa, and Nataniel Ruiz. Re- capture: Generative video camera controls for user-provided videos using masked video fine-tuning, 2024. 3

    2024

  37. [46]

    Monst3r: A simple approach for estimating geometry in the presence of motion, 2024

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jam- pani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming- Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion, 2024. 2, 3

    2024

  38. [47]

    I2vgen-xl: High-quality image-to-video synthe- sis via cascaded diffusion models, 2023

    Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jin- gren Zhou. I2vgen-xl: High-quality image-to-video synthe- sis via cascaded diffusion models, 2023. 3

    2023

  39. [48]

    Particlesfm: Exploiting dense point trajecto- ries for localizing moving cameras in the wild, 2022

    Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. Particlesfm: Exploiting dense point trajecto- ries for localizing moving cameras in the wild, 2022. 3

    2022

  40. [49]

    Cami2v: Camera-controlled image-to-video dif- fusion model, 2024

    Guangcong Zheng, Teng Li, Rui Jiang, Yehao Lu, Tao Wu, and Xi Li. Cami2v: Camera-controlled image-to-video dif- fusion model, 2024. 2, 3, 5, 6, 7, 8

    2024

  41. [50]

    Stereo magnification: Learning view syn- thesis using multiplane images, 2018

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view syn- thesis using multiplane images, 2018. 2, 5 10

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.