The authors of One Shot Any View present, in an arXiv paper, a system that turns a recorded clip into a reshot video of the same performance, following a camera trajectory the user specifies (arXiv). In the paper’s own Table 2, on 50 synchronized pairs, it has the lowest overall camera pose errors: RotErr 1.895 and TransErr 0.564, against 2.913 and 0.730 for Vista4D and 3.067 and 0.925 for CameraAnything (arXiv). The authors also say the system outperforms existing reshooting methods in both objective and human evaluations; these are their own figures, not an independent test.
What the system does
The task is reshooting: the input is one recorded take, and the output is the same performance seen along a camera path the user chooses. According to the paper, the generator produces 768P video natively, and an optional one-step diffusion super-resolution stage upscales it to 1440P. The authors say a geometry-anchored RoPE offset keeps the generated video in time with the source, while largely stopping the output from reproducing the degraded look of the point video. For real-world training data, the paper describes about 30k synchronized takes, each filmed by two to four cameras, from which about 10k two-camera clips of 3–15 s remain after selection.
Pose errors on 50 pairs
| Table block | Method | RotErr (deg) ↓ | TransErr ↓ | Source page |
|---|---|---|---|---|
| All pairs ( n=50 ) | Vista4D | 2.913 | 0.730 | https://arxiv.org/html/2611.24253 |
| All pairs ( n=50 ) | CameraAnything | 3.067 | 0.925 | https://arxiv.org/html/2611.24253 |
| All pairs ( n=50 ) | Ours | 1.895 | 0.564 | https://arxiv.org/html/2611.24253 |
| Small viewpoint change: pixel coverage >50% ( n=25 ) | Vista4D | 1.076 | 0.502 | https://arxiv.org/html/2611.24253 |
| Small viewpoint change: pixel coverage >50% ( n=25 ) | CameraAnything | 2.193 | 0.961 | https://arxiv.org/html/2611.24253 |
| Small viewpoint change: pixel coverage >50% ( n=25 ) | Ours | 1.580 | 0.333 | https://arxiv.org/html/2611.24253 |
| Large viewpoint change: pixel coverage <50% ( n=25 ) | Vista4D | 4.751 | 0.957 | https://arxiv.org/html/2611.24253 |
| Large viewpoint change: pixel coverage <50% ( n=25 ) | CameraAnything | 3.941 | 0.888 | https://arxiv.org/html/2611.24253 |
| Large viewpoint change: pixel coverage <50% ( n=25 ) | Ours | 2.210 | 0.795 | https://arxiv.org/html/2611.24253 |
Table 2 measures relative camera pose errors on synchronized video pairs, taking as the reference the poses estimated from the source frames and the recorded target frames. TransErr is given as a fraction of the median baseline of the reference clip, and the table reports means over each coverage group and over all 50 pairs, with each pair weighted equally. Across all pairs, One Shot Any View scores 1.895 RotErr and 0.564 TransErr, the lowest of the three methods on both. The picture changes with viewpoint: on small viewpoint changes, Vista4D has the lower RotErr, 1.076 against 1.580 for One Shot Any View. On large viewpoint changes, Vista4D’s RotErr rises to 4.751, while One Shot Any View rises to 2.210. CameraAnything scores 2.193 RotErr on small changes and 3.941 on large ones. The overall advantage in RotErr therefore comes from the large-change group, where the table shows One Shot Any View ahead of both baselines. On TransErr the table is more favorable: One Shot Any View has 0.333 on small changes and 0.795 on large ones, the lowest value in each group.
The authors explain the gap by saying Vista4D leans more heavily on the point video and tracks its positions closely, which helps while that video is reliable. Each group holds 25 pairs, so the group-level differences rest on small samples and come from the authors’ own evaluation.
Limits the authors state
The paper reports an identity drift in small faces: when a face covers only a few pixels in the native 768P output, the generator keeps little fine detail, and super-resolution can then produce sharper features that resemble a different person. This bears on the 1440P option, since the upscaling stage is where that drift appears. Another stated limit is unobserved content: if the source camera films a performer from behind and the requested view shows the face, a frontal reference photo taken at another time can guide face shape, but the model must still infer the expression at each moment. A generated expression may look natural yet differ from what a frontal camera would have recorded. Because each target view is generated independently, two views can show different expressions at the same moment.
The authors also note that training footage sometimes shows a performer outside the source camera’s field of view, so the generator occasionally adds an extra person when reshooting a clip. The current system processes prerecorded clips offline, and the authors treat real-time streaming as future work.
What to take from it
For practitioners, the main point is where the method helps: the paper’s numbers favor it when the viewpoint change is large and the point video degrades. The pose figures come from the authors’ own comparison with Vista4D and CameraAnything, so they show relative standing in that setup rather than an independent ranking.
What the sources we read do not state
We check six facts on every story in the full text of each source. The pages listed under Sources do not state a price or hardware requirements.
How we built this
Published resultsTable 2 of the paper: Relative camera pose errors on synchronized video pairs
Rows of Table 2 of the paper (arXiv HTML version), copied cell for cell by our code; a two-row header is shown as "group: column"; an empty name cell repeats the name above it. These are the authors' own figures, not an independent test. The first column is the block of the table the row sits in.
Sources
The sources this article was written from. Links open the original page.



