Filming a scene from an additional angle after production usually requires a costly reshoot. We present One Shot Any View, a system that takes a recorded clip and generates a reshot video of the same performance along a user-specified camera trajectory. As in recent geometry-grounded methods, the system conditions generation on the source video and an explicit 3D prior, namely a point video rendered along the requested trajectory from a feed-forward reconstruction of the source video. We target human-centric cinematic videos, which are often shot with long lenses, shallow depth of field, and low-key lighting, all of which degrade feed-forward reconstruction. Existing geometry-grounded methods tend to inherit the degraded point video, and their output quality drops visibly. Two designs make such point videos usable. First, in addition to rendered videos of 3D assets, we train on real-world multi-camera videos, whose point videos are produced by the same reconstruction pipeline as at inference. This training-inference consistency enables the generator to take only the composition of people and scene elements from the point video, and their appearance from the source video. Second, a geometry-anchored RoPE offset keeps the generated video temporally synchronized with the source while largely preventing the output from copying the degraded appearance of the point video. Meanwhile, since the generator is a pretrained reference-to-video (R2V) model, the system also accepts optional character and scene reference images, which supply the appearance of regions that the source never shows. In both objective and human evaluations, One Shot Any View outperforms existing reshooting methods.
Figures & tables
Figure 2: Reshooting from degraded target-view point videos. Each row shows the source frame, the point video frame at the requested viewpoint, and the generated frame after super-resolution. In (a), the target camera looks down on a person reclining outdoors. The point video stretches the person and the seat into streaks, while the generated view keeps the person and the seat in place as seen from above and fills the holes. In (b), moving the camera from a side view toward a frontal view leaves part of the face distorted and fragmented in the point video, while the generated view completes the face with high quality. In (c), only scattered points remain after a large viewpoint change in an office scene. The generated view shows the people around the desk from the elevated target viewpoint and keeps them in their original positions.
Rendered data
Real data
Inference input
Training stage
1 and 2
2
–
Geometry
ground truth
estimated
estimated
Image appearance
rendered
real
real
Point video quality
clean
degraded
degraded
Table 1: Training data and inference input compared. Estimated geometry comes from a feed-forward reconstruction model. Degraded point videos show deformed faces, lost detail, holes, and flicker (Figure 2 ), whereas rendered point videos have holes only where the source does not see the scene.
Figure 3: Rendered data. Each row shows a source frame, the point-video frame rendered into the target camera from the depth of the 3D engine, and the target frame. Wherever the source sees a surface, the point video is complete and aligned with the target. Black regions appear only where the source cannot see the scene.
Figure 4: Selecting real multi-camera data. The reconstruction of each frame pair is scored by its self-consistency, by its agreement with correspondences between the real images, and by the alignment of its rendering with the target frame. For each selected clip, we render the reconstruction of camera A into camera B at the original frame rate to obtain GA→B , and swap the cameras to obtain GB→A .
Figure 5: Real-world data. Each row shows a source frame, the point-video frame rendered into the target camera from feed-forward reconstruction, and the target frame. The point videos are sparser than those of rendered data, with holes along depth discontinuities and with stretched surfaces, yet people and scene elements land where the target recording shows them. The images shown here are AI-edited for illustration.
Figure 6: Video reshooting from a single source view. (a) The system estimates per-frame depth and camera poses and lifts source pixels into a colored point cloud. (b) Rendering the points along the requested camera trajectory yields a target-view point video. The R2V generator uses this video, the source clip, a structured caption, and optional character and scene images to generate the new view.
Figure 7: Conditional video generation with the R2V backbone. Prompt, visual, and target tokens form one packed sequence processed by a stack of Transformer blocks with joint self-attention and trainable LoRA adapters. The video head predicts the target velocity for iterative latent updates. The decoder maps the final latent to a 768P video.
Figure 8: One-step video super-resolution. (a) Schematic of a training pair in which the generated low-resolution video and the high-resolution target recording depict the same view at the same time. (b) A native 768P frame and its enhanced 1440P output are shown with crops of the same facial region.
Figure 9: Temporal RoPE layouts. Circles, squares, and diamonds mark source, point, and target frames, respectively. (a) Each stream occupies a separate temporal block. (b) Source, point, and target frames at each timestamp share the same RoPE coordinates. (c) The target coordinates shift by 0<δ<C/2 relative to the corresponding point frames, where C is the spacing between adjacent point frames.
Method
RotErr (deg) ↓
TransErr ↓
All pairs ( n=50 )
Vista4D [ 11 ]
2.913
0.730
CameraAnything [ 10 ]
3.067
0.925
Ours
1.895
0.564
Small viewpoint change: pixel coverage >50% ( n=25 )
Vista4D
1.076
0.502
Table 2: Relative camera pose errors on synchronized video pairs. Pose errors use the relative poses estimated from source and recorded target frames as the reference. TransErr is expressed as a fraction of the reference clip’s median baseline.
Comparison
Camera
Identity & app.
Temporal
Quality
Overall
Ours vs. Vista4D
4%
18%
28%
22%
38%
Ours vs. CameraAnything
72%
68%
36%
68%
92%
Table 3: User preference (GSB). Each cell reports the net GSB score (G−B)/(G+S+B) as a percentage for our method against the baseline. Positive values indicate a preference for our method.
Figure 10: Comparison with Vista4D and CameraAnything. Each row shows the source, the point video, and the three generated views at the same timestamp. Frames run from top to bottom within each case. All methods use the same source clips and target trajectories.
Figure 11: Effect of real multi-camera training data. Both variants reshoot the same source frame from the viewpoint of the point video. Real training footage reduces the transfer of point-video artifacts to the dancers (top) and to the attendant and the objects on the cart (bottom).
Figure 12: Effect of the temporal RoPE layout. Each row shows the same instant across views. Temporal isolation changes when the player’s foot meets the ball. Co-registration superimposes two copies of the player. The geometry-anchored offset preserves both the timing of the touch and a single image of the player.
Figure 13: Effect of reference images. (a) The scene reference supplies the appearance of the courtyard beyond the arcade. (b) The source shows the runners from behind. Boxes mark the face and upper chest of the runner in blue, in his reference portrait and in both target views. The enlargements show the shirt logo and its location relative to the collar.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 14: Additional reshooting results. Each case shows the reconstructed point cloud and the source and target cameras on the left. On the right, source frames (top) and generated target-view frames (bottom) are paired at matching times, progressing from left to right. Cyan marks the source camera and orange marks the target camera.
Precise camera control for reshooting dynamic videos is bottlenecked by the severe scarcity of paired multi-view data for non-rigid scenes. We overcome this limitation with a highly scalable self-supervised framework capable of leveraging internet-scale monocular videos. Our core contribution is the generation of pseudo multi-view training triplets, consisting of a source video, a geometric anchor, and a target video. We achieve this by extracting distinct smooth random-walk crop trajectories from a single input video to serve as the source and target views. The anchor is synthetically generated by forward-warping the first frame of the source with a dense tracking field, which effectively simulates the distorted point-cloud inputs expected at inference. Because our independent cropping strategy introduces spatial misalignment and artificial occlusions, the model cannot simply copy information from the current source frame. Instead, it is forced to implicitly learn 4D spatiotemporal structures by actively routing and re-projecting missing high-fidelity textures across distinct times and viewpoints from the source video to reconstruct the target. At inference, our minimally adapted diffusion transformer utilizes a 4D point-cloud derived anchor to achieve state-of-the-art temporal consistency, robust camera control, and high-fidelity novel view synthesis on complex dynamic scenes.
Re-rendering an existing video from a novel camera viewpoint requires the output to follow the prescribed camera trajectory while preserving the appearance and dynamics of the original scene across every frame. Existing methods rely on per-frame pose embeddings, noisy point-cloud renderings, or implicit learned correspondences, none of which provides an explicit, temporally continuous link between source and target pixels. We propose Track2View, which conditions a video diffusion transformer on paired 3D point tracks: sparse trajectories of scene points projected into both the source and target camera views. These tracks provide explicit spatiotemporal correspondences that are temporally continuous by construction, encoding what content should appear where and when. At the core of Track2View is a dual-view track conditioner that transfers visual context from source to target view through parameter-free geometric operations and learned temporal aggregation, ensuring generalization to arbitrary camera trajectories without memorizing specific motions. We further introduce a data curation pipeline that extracts one-to-one track correspondences by running a 3D point tracker on temporally concatenated multi-camera view pairs. On a 400-video benchmark spanning static and dynamic scenes, Track2View achieves state-of-the-art results across visual quality, view synchronization, and camera accuracy, reducing rotation error by 30-65% and translation error by 61-72% relative to leading baselines. Project page is available at this https URL: https://qjizhi.github.io/track2view
Feng Qiao, Zhaochong An, Zhexiao Xiong +2
Washington University in St. Louis · University of Copenhagen
We present Vista4D, a robust and flexible video reshooting framework that grounds the input video and target cameras in a 4D point cloud. Specifically, given an input video, our method re-synthesizes the scene with the same dynamics from a different camera trajectory and viewpoint. Existing video reshooting methods often struggle with depth estimation artifacts of real-world dynamic videos, while also failing to preserve content appearance and failing to maintain precise camera control for challenging new trajectories. We build a 4D-grounded point cloud representation with static pixel segmentation and 4D reconstruction to explicitly preserve seen content and provide rich camera signals, and we train with reconstructed multiview dynamic data for robustness against point cloud artifacts during real-world inference. Our results demonstrate improved 4D consistency, camera control, and visual quality compared to state-of-the-art baselines under a variety of videos and camera paths. Moreover, our method generalizes to real-world applications such as dynamic scene expansion and 4D scene recomposition. See our project page for results, code, and models: https://eyeline-labs.github.io/Vista4D