Pith. sign in

REVIEW 4 major objections 7 minor 41 references

One video model can reshoot a clip with new viewpoint, focal length, resolution, and shot cuts in a single pass.

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 · grok-4.5

2026-07-31 11:02 UTC pith:LDMCL76I

load-bearing objection Real unification of multi-shot retakes with focal/resolution control, but the intrinsic story is mostly synthetic crops and the real-data edge is thin. the 4 major comments →

arxiv 2607.24591 v1 pith:LDMCL76I submitted 2026-07-27 cs.CV

CameraAnything: Refilming Videos with Arbitrary Camera Control

Yixuan Li , Yanhong Zeng , Ka Leong Cheng , Jiayi Zhu , Hanlin Wang , Wen Wang , Yihao Meng , Hao Ouyang
show 6 more authors
Qiuyu Wang Yue Yu ZiDong Wang Yiyuan Zhang Yujun Shen Dahua Lin
This is my paper
classification cs.CV
keywords camera-controlled video editingPlücker rays3D RoPEfocal length controlmulti-shot transitionsnative resolution editingvideo diffusion transformersynthetic multi-camera data
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.

CameraAnything claims to be the first unified system that refilms an existing video under full camera control: new trajectories, multi-shot cuts, focal-length changes, and native resolution or aspect-ratio changes, all in one generation. Prior tools either rebuild expensive 3D scenes or only move the camera while locking lens and frame size, so they cannot reproduce cinematic moves like dolly zooms or reformat a take for phone versus TV. The authors condition a video diffusion transformer with per-pixel Plücker rays and resolution-aware 3D rotary encodings tied to the target frame grid, and train on synchronized multi-camera synthetic clips with an orthogonal sampling strategy that mixes extrinsic, focal, and resolution edits. If the approach holds, ordinary footage can be re-shot with expressive camera language without multi-camera rigs or full 3D reconstruction, which matters for cinema-style editing and cross-platform content adaptation.

Core claim

A single latent video diffusion model can jointly obey target extrinsic pose (including abrupt multi-shot cuts), focal length, and native output resolution when both camera conditioning and spatial positional encoding are built from per-pixel Plücker rays and resolution-aware 3D RoPE on the target latent, trained with orthogonally sampled synchronized synthetic multi-camera pairs.

What carries the argument

Per-pixel Plücker ray injection into resolution-aware 3D RoPE in self-attention: each target token gets a 6-DoF ray (origin and direction) that encodes pose and focal length, applied as a phase shift alongside RoPE whose spatial coordinates come from the target latent grid, so viewpoint, zoom, and resolution change together without cropping or outpainting.

Load-bearing premise

Geometry and appearance learned almost entirely from synthetic multi-camera scenes with animated characters transfer well enough to real videos—even when source cameras are noisy estimates—that the model stays geometrically faithful.

What would settle it

On real videos with independently measured ground-truth target cameras and lenses, check whether generated frames match the commanded pose, focal length, and resolution (rotation/translation error and reconstruction metrics) as well as they do on held-out synthetic scenes; a large real-world drop would falsify the transfer claim.

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

If this is right

  • A continuous single-take can be turned into a multi-shot edit with specified cuts inside one generation pass.
  • Focal length and camera path can be changed together, enabling effects such as dolly zoom without separate optical and motion stages.
  • The same source can be re-rendered at native portrait, landscape, or square resolutions without crop-or-outpaint pipelines.
  • Cinematic retakes and cross-device reformatting become feasible from ordinary single-camera footage rather than multi-camera capture or full 3D rebuilds.

Where Pith is reading between the lines

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

  • If Plücker-on-target-RoPE is the binding constraint, the same injection pattern may extend to other intrinsics (aperture, shutter) once paired data exist.
  • Heavy reliance on synthetic multi-view pairs suggests that real multi-camera or light-field corpora would be the highest-leverage next training signal for closing residual real-world geometry gaps.
  • Native resolution control via target-grid RoPE could generalize to other video editors that today still fix aspect ratio at training time.

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

4 major / 7 minor

Summary. CameraAnything fine-tunes Wan2.1-T2V-1.3B for video-to-video camera editing by concatenating source and target tokens, representing cameras with per-pixel Plücker rays, injecting them through zero-initialized RoPE phase shifts, and recomputing 3D-RoPE coordinates for target-resolution latent grids. A structured Unreal Engine/BEDLAM dataset supplies synchronized single-shot, multi-shot, and re-focal clips, while independent sampling of extrinsic, focal, and aspect-ratio edits is intended to disentangle the three control axes. The reported results improve synthetic reconstruction and ViPE-estimated camera errors over prior extrinsic-only methods and show qualitative multi-shot and resolution adaptation. The extrinsic and synthetic focal results are credible, but the central unification claim is not yet adequately tested on real intrinsic control or on resolution changes requiring content beyond the source frustum.

Significance. If validated, a single generative editor supporting viewpoint changes, cinematic cuts, focal changes, and output-format adaptation would be useful for video post-production and would advance beyond extrinsic-only systems such as ReCamMaster. The paper has real strengths: a structured paired-data protocol, explicit re-focal pairs with identical trajectories, an orthogonal multi-task training design, representation/injection ablations, per-configuration synthetic metrics, external ViPE-based pose evaluation on real footage, and a human-preference study. These make the extrinsic-control contribution and the synthetic focal supervision reasonably persuasive. However, the evidence presently supports a narrower claim than “arbitrary” joint intrinsic/extrinsic control: real focal adherence is unmeasured, and the main resolution evaluation appears to test crop-like recomposition rather than generation outside the source view.

major comments (4)
  1. [§3, §5.1, Table 6] All raw footage is rendered at 832×832 and aspect variants are produced by parameterized center crops. The sole synthetic resolution configuration is described as the source “re-encoded at portrait resolution (832×480)”; Table 6 gives it PSNR 20.54, close to focal-only (21.78) and far above extrinsic edits (~14.5). Thus, as described, the benchmark does not establish editing where the target frustum contains content absent from the source, despite the claims “without cropping or outpainting” and cross-platform adaptation in Fig. 1. Please add crop/recompose versus FOV-expanding or cross-aspect cases, including evaluation of newly visible regions, or narrow the claim.
  2. [§5.1–5.2, Tables 2 and 5] The real-world evaluation contains no focal or FOV adherence metric: RotErr/TransErr measure only extrinsics, VBench is generic perceptual quality, and the user study visualizes target pose rather than explicitly scoring focal alignment. This is load-bearing because focal control is one of the paper’s two main intrinsic-control novelties, while §5.1 trains sources only from the fixed-focal subset. Please report real focal/FOV adherence using estimated intrinsics or a controlled human/geometric protocol, with uncertainty and sensitivity to source-camera estimation errors.
  3. [§5.1, Tables 1 and 6] The “Ours” PSNR of 15.88 appears to average the five non-resolution configurations in Table 6, including focal-only and extrinsic+focal cases. ReCamMaster supports only extrinsic conditioning, yet the manuscript does not state what intrinsic parameters it receives in those cases. The aggregate comparison therefore conflates missing focal-control capability with camera-following quality. Likewise, outpainting baselines are compared on a crop-like resolution case without specifying their masks or instructions. Please report per-configuration baseline results and exact conditioning, and use capability-matched subsets for claims of superiority.
  4. [Tables 1–3] No sample-level variance or paired significance is reported. This matters because the real resolution differences are tiny (e.g., Table 2 Aesthetic 0.5198 versus 0.5166 and Imaging 0.6732 versus 0.6728), while the selected Plücker+RoPE variant in Table 3 improves PSNR/SSIM over AdaLN only slightly and has worse LPIPS (0.372 versus 0.368). Please provide per-video standard deviations or bootstrap confidence intervals and paired tests, and qualify “best reconstruction fidelity” and “consistently well” accordingly.
minor comments (7)
  1. [§5.1] “Source portrait” is paired with the resolution 832×480, which is landscape under the usual width×height convention. Please clarify the resolution ordering and intended aspect-ratio terminology.
  2. [§4.3, Table 4] The term “identity samples” is not precisely defined. §4.3 says an all-identity draw is promoted to a focal-change sample, whereas Table 4 suggests identity samples are important for decoupling. Please state exactly which dimensions are held fixed in these samples.
  3. [§4.2, Eq. (6), Fig. 3] Clarify whether Plücker-conditioned phase shifts are applied to source tokens, target tokens, or both, and how their relative phases are composed in self-attention. Figure 3 suggests both cameras are used, while the surrounding prose emphasizes the target latent.
  4. [§4.1–4.2, Eqs. (1) and (3)] Specify the mapping from pixel-space Plücker rays to VAE latent tokens after spatial stride and patchification. Also note the principal-point convention in Eq. (1); many geometry pipelines use (W−1)/2 and (H−1)/2 rather than W/2 and H/2.
  5. [Appendix B, Table 5] The user study should report question wording, randomization, tie handling, confidence intervals, and whether focal-control cases were shown to a baseline that cannot accept focal instructions.
  6. [§3, §5.1, Appendix A] Please provide release status and sufficient details for reproducing the dataset and training: focal/FOV distributions, crop parameterization, trajectory sampler, train/test scene split, sampling seeds, inference settings, and whether code or rendered data will be available.
  7. [Appendix A, Figs. 5–9] “Our dataset have about 300,000 videos” should be “has”; the large figure legends would also benefit from clearer trajectory labels and explicit indication of which regions are newly synthesized.

Circularity Check

0 steps flagged

No circularity: standard supervised video-to-video training with held-out metrics, not predictions forced by definition or self-citation.

full rationale

CameraAnything is an engineering/ML systems paper, not a first-principles derivation. The load-bearing chain is: (1) represent target cameras as per-pixel Plücker rays and inject them via RoPE phase shifts into a frozen Wan2.1 DiT, with resolution-aware 3D RoPE on the target latent grid; (2) train on paired source/target clips from a self-built Unreal multi-camera corpus under an orthogonal sampling schedule over extrinsics, focal length, and aspect ratio; (3) evaluate reconstruction (PSNR/SSIM/LPIPS), camera error (ViPE RotErr/TransErr), VBench, and human preference on held-out synthetic scenes and real videos. None of these steps reduces a claimed prediction to its own fitted input or definition. Ablations (Tables 3–4) compare representation/injection and training variants empirically; baselines (ReCamMaster, TrajectoryCrafter, Follow-Your-Canvas, VACE) are external. Self-built synthetic data and citations to prior camera encodings (Plücker, RoCE) are normal methodological choices, not uniqueness theorems or self-citation chains that force the result. Concerns that resolution supervision is mostly center-crops of square renders, or that real-world focal adherence is unmeasured, are evaluation-validity issues—not circularity under the stated patterns. Score 0; steps empty.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 2 invented entities

As an empirical systems paper, load-bearing premises are domain assumptions of pinhole cameras, transfer from synthetic multi-view renders, and the adequacy of Plücker+RoPE conditioning inside a frozen pretrained video DiT—not new physical entities. Free parameters are ordinary ML hyperparameters and hand-set task-sampling rates rather than constants fitted to claim a law.

free parameters (4)
  • Orthogonal task sampling probabilities = 0.2/0.4/0.4; p_focal=0.5; p_aspect=0.2
    Extrinsics type probs 0.2/0.4/0.4; focal change 0.5; aspect change 0.2; forced non-identity rule (§4.3). Chosen by authors to shape multi-task learning; not derived.
  • Fine-tuning hyperparameters = 20k steps, 5e-5, bs=32
    20k steps, lr 5e-5, batch 32, trainable subset (self-attn, camera encoder/adapter), zero-init adapters (§5.1).
  • Resolution training buckets = {480×832,...,320×320} listed set
    Discrete set of HxW pairs used for native resolution training (§5.1).
  • Synthetic focal set / FOV sampling range = 18–50mm-equivalent patterns
    Base focals from {18,24,35,50}mm-equivalent; Group C varies FOV in 18–50mm range with zoom patterns (App. A).
axioms (6)
  • domain assumption Pinhole camera model with intrinsics from horizontal FOV and extrinsics as c2w matrices, poses normalized to first source frame.
    §4.1 Eqs. (1)–(2); standard CV camera model underlying Plücker construction.
  • domain assumption Per-pixel Plücker rays uniquely encode the joint effect of extrinsics and intrinsics for each token and adapt when resolution/FOV change.
    §4.2 Eqs. (3); justified by classical ray geometry and prior camera-conditioned generators.
  • ad hoc to paper Injecting camera geometry as RoPE phase shifts (RoCE-style) steers attention without destroying the pretrained generative prior when zero-initialized.
    §4.2 Eq. (6) and Design Choice; supported by their ablation but not a theorem.
  • domain assumption Synchronized UE5 multi-camera renders with BEDLAM humans are sufficient supervision for real-world camera-controlled editing.
    §3 and App. A dataset construction; central transfer assumption for real eval.
  • domain assumption Frame-dimension concatenation of source and noisy target latents enables content-consistent video-to-video generation in Wan2.1 DiT.
    §4.1 following ReCamMaster; architectural prior from cited work.
  • domain assumption ViPE-estimated poses are accurate enough to score camera following on real videos.
    §5.1 real-world metrics RotErr/TransErr depend on this external estimator.
invented entities (2)
  • CameraAnything orthogonal multi-task training protocol (independent extrinsic/focal/resolution sampling with identity promotion) no independent evidence
    purpose: Disentangle and compose camera control dimensions despite coupled appearance effects.
    Described in §4.3; a training procedure, not a physical entity. Evidence is internal ablations (Table 4), not an external falsifiable object.
  • Structured 20-clip UE multi-camera scene package (Groups A/B/C including re-focal and multi-shot) no independent evidence
    purpose: Supply paired supervision for joint intrinsic/extrinsic and multi-shot editing.
    §3/App. A dataset design. Independent evidence would require public release and third-party reuse; not indicated in the text.

pith-pipeline@v1.2.0-grok45-kimik3 · 19691 in / 4056 out tokens · 76735 ms · 2026-07-31T11:02:32.184629+00:00 · methodology

0 comments
read the original abstract

We introduce CameraAnything, the first unified framework for camera controlled video editing that enables joint control of both intrinsic and extrinsic camera parameters. Existing approaches either rely on expensive 3D reconstruction to achieve full camera functionality or restrict editing to extrinsic parameter manipulation. Moreover, the coupled influence of intrinsic and extrinsic parameters on video appearance makes disentangled modeling particularly challenging. To address this, we adopt per-pixel Pl\"ucker ray injection alongside resolution-aware 3D RoPE in self-attention, building both camera conditioning and spatial positional encoding on the target latent to jointly control camera position, focal length, and native resolution editing without cropping or outpainting. To overcome the scarcity of paired training data, we further develop a scalable synthetic pipeline that constructs diverse dynamic scenes through structured multi-camera recording and generates synchronized videos with varied camera configurations. With a tailored orthogonal training strategy, CameraAnything enables expressive video reshooting with arbitrary viewpoint control, focal length adjustment, resolution adaptation, and multi-shot transitions within a single generation process, offering strong practical value for cinematic video editing and cross-platform content adaptation in video production.

Figures

Figures reproduced from arXiv: 2607.24591 by Dahua Lin, Hanlin Wang, Hao Ouyang, Jiayi Zhu, Ka Leong Cheng, Qiuyu Wang, Wen Wang, Yanhong Zeng, Yihao Meng, Yixuan Li, Yiyuan Zhang, Yue Yu, Yujun Shen, Zidong Wang.

Figure 1
Figure 1. Figure 1: We propose CameraAnything, the first unified video editing framework that enables comprehensive camera control, including novel trajectory manipulation, multi-shot transitions, resolution adaptation, and focal length adjustment in a single framework, demonstrating strong potential for cinema-level video editing. arXiv:2607.24591v1 [cs.CV] 27 Jul 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dataset Example. The red box represents single-shot settings with consistent camera extrinsics, including both fixed and varying focal lengths. The orange and blue boxes each represent a different shot in multi-shot sequences. We construct diverse 3D environments populated with animated human char￾acters to simulate real-world dynamic scenes. To enable the model to capture the spatial correlations between … view at source ↗
Figure 2
Figure 2. Figure 2: Additional dataset details are provided in Appendix A. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of CameraAnything. CameraAnything is built upon Wan2.1- T2V-1.3B [33], which includes a 3D VAE for video latent encoding and a diffusion Transformer composed of multiple DiT blocks. Given a source video together with its camera parameters and a target camera specification, CameraAnything represents cameras using Plücker rays to provide accurate per-token geometric modeling, which is injected into … view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of three camera representation and injection mechanisms. (a) Token-wise Addition: Pi ∈ R 3×4 and Ki ∈ R 3×3 are flattened and concatenated into ci ∈ R 21, then linearly encoded and added to all tokens. (b) AdaLN Injection: per-pixel camera embeddings based on the Plücker ray representation are encoded by a learnable linear layer to generate scale and shift parameters that modulate the Layer No… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on real-world videos with TrajectoryCrafter [37] and ReCamMaster [2]. Given the same source video and target retake camera parameters, our method demonstrates more accurate camera trajectory following under challeng￾ing scenarios, including large viewpoint changes and multi-cut shot transitions, while maintaining higher visual fidelity and geometric consistency. Source Video Target V… view at source ↗
Figure 6
Figure 6. Figure 6: In-the-wild results. Our method generalizes to diverse real-world scenes despite being trained on synthetic data only. than feature-space modulation. We include qualitative comparisons for different injection designs in [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparisons for the ablation study. Given the same source video and target retake camera parameters, Plücker ray representation with RoPE-based injection yields stronger geometric consistency, with fewer artifacts (highlighted by red dashed boxes) compared to other variants. Green dashed boxes indicate the ground truth [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Results under large viewing-angle and resolution changes. Each row shows a source video (left) and our result (right). Source Target Source Target [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Results on two movie cases. Each row shows a source frame (left) and our edited result (right), illustrating generalization to cinematic footage beyond synthetic training data [PITH_FULL_IMAGE:figures/full_fig_p022_9.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

41 extracted references · 10 linked inside Pith

  1. [1]

    Adobe: Dolly zoom shot: Definition and examples.https://www.adobe.com/ sg / creativecloud / video / production / cinematography / camera - shots - and - angles/dolly-zoom-shot.html(2026), accessed: 2026-06-30

    2026

  2. [2]

    In: ICCV (2025)

    Bai, J., Xia, M., Fu, X., Wang, X., Mu, L., Cao, J., Liu, Z., Hu, H., Bai, X., Wan, P., et al.: Recammaster: Camera-controlled generative rendering from a single video. In: ICCV (2025)

    2025

  3. [3]

    In: ICLR (2025)

    Bai, J., Xia, M., Wang, X., Yuan, Z., Liu, Z., Hu, H., Wan, P., ZHANG, D.: Syn- cammaster: Synchronizing multi-camera video generation from diverse viewpoints. In: ICLR (2025)

    2025

  4. [4]

    In: CVPR

    Black, M.J., Patel, P., Tesch, J., Yang, J.: BEDLAM: A synthetic dataset of bod- ies exhibiting detailed lifelike animated motion. In: CVPR. pp. 8726–8737. IEEE (2023)

    2023

  5. [5]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Chen, Q., Ma, Y., Wang, H., Yuan, J., Zhao, W., Tian, Q., Wang, H., Min, S., Chen, Q., Liu, W.: Infinite-canvas: Higher-resolution video outpainting with ex- tensive content generation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 2150–2158 (2025)

    2025

  6. [6]

    arXiv preprint arXiv:2511.17185 (2025)

    Chen, Y., Ye, Z., Fang, Z., Chen, X., Zhang, X., Liu, J., Wang, N., Liu, H., Zhang, G.: Postcam: Camera-controllable novel-view video generation with query-shared cross-attention. arXiv preprint arXiv:2511.17185 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    Epic Games: Unreal Engine.https://www.unrealengine.com/(2023), accessed: 2026-06-30

    2023

  8. [8]

    arXiv preprint arXiv:2601.05239 (2026)

    Fu, X., Tang, S., Shi, M., Liu, X., Gu, J., Liu, M.Y., Lin, D., Lin, C.H.: Plenoptic video generation. arXiv preprint arXiv:2601.05239 (2026)

    arXiv 2026

  9. [9]

    Gao, R., Hołyński, A., Henzler, P., Brussee, A., Martin-Brualla, R., Srinivasan, P., Barron, J.T., Poole, B.: Cat3d: create anything in 3d with multi-view diffusion models.In:Proceedingsofthe38thInternationalConferenceonNeuralInformation Processing Systems. pp. 75468–75494 (2024)

    2024

  10. [10]

    In: ICLR (2025)

    He, H., Xu, Y., Guo, Y., Wetzstein, G., Dai, B., Li, H., Yang, C.: Cameractrl: Enabling camera control for text-to-video generation. In: ICLR (2025)

    2025

  11. [11]

    In: ICCV (2025)

    He, H., Yang, C., Lin, S., Xu, Y., Wei, M., Gui, L., Zhao, Q., Wetzstein, G., Jiang, L., Li, H.: Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models. In: ICCV (2025)

    2025

  12. [12]

    arXiv preprint arXiv:2512.05672 (2025)

    Hong, Y., Lee, S., Chung, H., Ye, J.C.: Inversecrafter: Efficient video recapture as a latent domain inverse problem. arXiv preprint arXiv:2512.05672 (2025)

    Pith/arXiv arXiv 2025

  13. [13]

    arXiv preprint arXiv:2506.05554 (2025)

    Hu, T., Peng, H., Liu, X., Ma, Y.: Ex-4d: Extreme viewpoint 4d video synthesis via depth watertight mesh. arXiv preprint arXiv:2506.05554 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    CoRRabs/2508.10934(2025)

    Huang, J., Zhou, Q., Rabeti, H., Korovko, A., Ling, H., Ren, X., Shen, T., Gao, J., Slepichev, D., Lin, C., Ren, J., Xie, K., Biswas, J., Leal-Taixé, L., Fidler, S.: Vipe: Video pose engine for 3d geometric perception. CoRRabs/2508.10934(2025)

    Pith/arXiv arXiv 2025

  15. [15]

    arXiv preprint arXiv:2512.25075 (2025)

    Huang, Z., Jeong, H., Chen, X., Gryaditskaya, Y., Wang, T.Y., Lasenby, J., Huang, C.H.: Spacetimepilot: Generative rendering of dynamic scenes across space and time. arXiv preprint arXiv:2512.25075 (2025)

    arXiv 2025

  16. [16]

    In: CVPR (2024)

    Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., et al.: Vbench: Comprehensive benchmark suite for video genera- tive models. In: CVPR (2024)

    2024

  17. [17]

    arXiv preprint (2025) CameraAnything: Refilming Videos with Arbitrary Camera Control 17

    HunyuanWorld, T.: Hunyuanworld 1.0: Generating immersive, explorable, and in- teractive 3d worlds from words or pixels. arXiv preprint (2025) CameraAnything: Refilming Videos with Arbitrary Camera Control 17

    2025

  18. [18]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Jiang, Z., Han, Z., Mao, C., Zhang, J., Pan, Y., Liu, Y.: Vace: All-in-one video creation and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17191–17202 (2025)

    2025

  19. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Kong,X.,Liu,S.,Lyu,X.,Taher,M.,Qi,X.,Davison,A.J.:Eschernet:Agenerative model for scalable view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9503–9513 (2024)

    2024

  20. [20]

    Advances in Neural Information Processing Systems (2025)

    Li, R., Yi, B., Liu, J., Gao, H., Ma, Y., Kanazawa, A.: Cameras as relative posi- tional encoding. Advances in Neural Information Processing Systems (2025)

    2025

  21. [21]

    CoRRabs/2504.15376(2025)

    Lin, Z., Cen, S., Jiang, D., Karhade, J., Wang, H., Mitra, C., Ling, T., Huang, Y., Liu, S., Chen, M., Zawar, R., Bai, X., Du, Y., Gan, C., Ramanan, D.: Towards understanding camera motions in any video. CoRRabs/2504.15376(2025)

    Pith/arXiv arXiv 2025

  22. [22]

    arXiv preprint arXiv:2512.05115 (2025)

    Liu, T., Chen, Z., Huang, Z., Xu, S., Zhang, S., Ye, C., Li, B., Cao, Z., Li, W., Zhao, H., et al.: Light-x: Generative 4d video rendering with camera and illumination control. arXiv preprint arXiv:2512.05115 (2025)

    arXiv 2025

  23. [23]

    In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers

    Luo, Y., Shi, X., Bai, J., Xia, M., Xue, T., Wang, X., Wan, P., Zhang, D., Gai, K.: Camclonemaster: Enabling reference-based camera control for video generation. In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers. pp. 1–10 (2025)

    2025

  24. [24]

    In: The Twelfth International Conference on Learning Representations (2024)

    Miyato, T., Jaeger, B., Welling, M., Geiger, A.: Gta: A geometry-aware attention mechanism for multi-view transformers. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  25. [25]

    arXiv preprint arXiv:2511.19827 (2025)

    Park, B., Kim, B.H., Chung, H., Ye, J.C.: Redirector: Creating any-length video retakes with rotary camera encoding. arXiv preprint arXiv:2511.19827 (2025)

    arXiv 2025

  26. [26]

    CoRR abs/1704.00675(2017)

    Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Gool, L.V.: The 2017 DAVIS challenge on video object segmentation. CoRR abs/1704.00675(2017)

    Pith/arXiv arXiv 2017

  27. [27]

    arXiv preprint arXiv:2506.13697 (2025)

    Seo, J., Han, J., Jung, J., Jin, S., Lee, J., Narihira, T., Fukuda, K., Shibuya, T., Ahn, D., Hu, S., et al.: Vid-camedit: Video camera trajectory editing with generative rendering from estimated geometry. arXiv preprint arXiv:2506.13697 (2025)

    Pith/arXiv arXiv 2025

  28. [28]

    Advances in Neural Information Processing Systems34, 19313–19325 (2021)

    Sitzmann, V., Rezchikov, S., Freeman, B., Tenenbaum, J., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems34, 19313–19325 (2021)

    2021

  29. [29]

    Song, C., Yang, Y., Zhao, T., Li, R., Zhang, C.: Worldforge: Unlocking emergent 3d/4dgenerationinvideodiffusionmodelviatraining-freeguidance.arXivpreprint arXiv:2509.15130 (2025)

    arXiv 2025

  30. [30]

    Neurocomputing568, 127063 (2024)

    Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced trans- former with rotary position embedding. Neurocomputing568, 127063 (2024)

    2024

  31. [31]

    arXiv preprint arXiv:2511.12921 (2025)

    Sun, H., Shen, L., Peng, Z., Wang, K., Wu, S., Zang, Y., Liu, T., Huang, Z., Zeng, X., Cao, Z., et al.: Generative photographic control for scene-consistent video cinematic editing. arXiv preprint arXiv:2511.12921 (2025)

    arXiv 2025

  32. [32]

    arXiv preprint arXiv:1812.01717 (2019)

    Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Fvd: A new metric for video generation. arXiv preprint arXiv:1812.01717 (2019)

    Pith/arXiv arXiv 2019

  33. [33]

    arXiv preprint arXiv:2503.20314 (2025)

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., et al.: Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025)

    Pith/arXiv arXiv 2025

  34. [34]

    arXiv preprint arXiv:2512.05076 (2025)

    Wang, Y., Zhang, Q., Cai, S., Wu, T., Ackermann, J., Kuang, Z., Zheng, Y., Rajič, F., Tang, S., Wetzstein, G.: Bullettime: Decoupled control of time and camera pose for video generation. arXiv preprint arXiv:2512.05076 (2025)

    arXiv 2025

  35. [35]

    arXiv preprint arXiv:2512.10286 (2025) 18 Yixuan Li et al

    Wu, X., Chen, X., Wang, Y., Qiao, Y.: Shotdirector: Directorially controllable multi-shot video generation with cinematographic transitions. arXiv preprint arXiv:2512.10286 (2025) 18 Yixuan Li et al

    arXiv 2025

  36. [36]

    arXiv preprint arXiv:2511.21024 (2025)

    Yang, Q., Yang, Y., Zeng, Y., Hu, X., Li, B., Yue, H., Yang, J., Jiang, P.T.: Cam- eramaster: Unified camera semantic-parameter control for photography retouching. arXiv preprint arXiv:2511.21024 (2025)

    arXiv 2025

  37. [37]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Yu, M., Hu, W., Xing, J., Shan, Y.: Trajectorycrafter: Redirecting camera trajec- tory for monocular videos via diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 100–111 (October 2025)

    2025

  38. [38]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Yuan, Y., Wang, X., Sheng, Y., Chennuri, P., Zhang, X., Chan, S.: Generative photography: Scene-consistent camera control for realistic text-to-image synthesis. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 7920–7930 (2025)

    2025

  39. [39]

    arXiv preprint arXiv:2512.07237 (2025)

    Zhang, C., Li, B., Wei, M., Cao, Y.P., Gambardella, C.C., Phung, D., Cai, J.: Unified camera positional encoding for controlled video generation. arXiv preprint arXiv:2512.07237 (2025)

    arXiv 2025

  40. [40]

    In: CVPR (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)

    2018

  41. [41]

    Zhong, L., Li, F., Huang, Y., Liu, J., Pei, R., Song, F.: Outdreamer: Video out- painting with a diffusion transformer. arXiv preprint arXiv:2506.22298 (2025) CameraAnything: Refilming Videos with Arbitrary Camera Control 19 CameraAnything: Refilming Videos with Arbitrary Camera Control Supplementary Material A Dataset Details Scene construction.Our datas...

    Pith/arXiv arXiv 2025