Video (camera) trajectory editing aims to synthesize new videos that follow user-defined camera paths while preserving scene content and plausibly inpainting previously unseen regions, upgrading amateur footage into professionally styled videos. Existing VTE methods struggle with precise camera control and long-range consistency because they either inject target poses through a limited-capacity embedding or rely on single-frame warping with only implicit cross-frame aggregation in video diffusion models. To address these issues, we introduce a new VTE framework that 1) explicitly aggregates information across the entire source video via a hybrid warping scheme. Specifically, static regions are progressively fused into a world cache then rendered to target camera poses, while dynamic regions are directly warped; their fusion yields globally consistent coarse frames that guide refinement. 2) processes video segments jointly with their history via a history-guided autoregressive diffusion model, while the world cache is incrementally updated to reinforce already inpainted content, enabling long-term temporal coherence. Finally, we present iPhone-PTZ, a new VTE benchmark with diverse camera motions and large trajectory variations, and achieve state-of-the-art performance with fewer parameters.
Figures & tables
Figure 2: Visual comparison between per-frame warping (a), our hybrid warping (b), and ground truth (c). Hybrid warping tends to produce more complete and source-aligned coarse frames.
Figure 3: Overview of our framework . Left: The hybrid warping scheme leverages the entire source video to construct coarse frames by processing dynamic and static regions separately, providing a global reference of the original scene content. Right: The CCDM conditions the generation on the coarse video via ControlNet, while source-frame tokens are concatenated with target tokens as inputs to the base T2V model to provide reliable motion and appearance priors.
Figure 4: Illustration of history-guided autoregressive generation . In each iteration, T⋆ previously generated frames serve as history to guide the synthesis of the next T frames, along with the corresponding T⋆+T source frames as input to produce the coarse frames and provide original scene context.
Figure 5: Illustration of progressive world cache update . Whenever a new segment is generated, we evenly sample C frames as anchors, where the newly inpainted regions are merged into the world cache. The updated regions are highlighted in red.
Figure 6: Illustration of VGGT depth vs. our corrected depths . The last two columns visualize warping results from the source to the target view using VGGT’s estimated depth and our corrected depth, respectively. VGGT’s depth leads to a poor warping result (third column), whereas our adjusted depth yields a more aligned outcome (last column).
Method
# Params
iPhone [ 9 ]
iPhone-PTZ
PSNR ↑
LPIPS ↓
FID ↓
PSNR ↑
LPIPS ↓
FID ↓
RecamMaster [ 2 ]
1.3B
10.73 / -
0.7830 / -
195.24 / -
11.64 / -
0.6981 / -
117.77 / -
TrajectoryCrafter [ 33 ]
5.3B
13.00 / -
0.6197 / -
145.58 / -
12.56 / -
0.5303 / -
105.30 / -
Gen3C [ 24 ]
6.7B
13.29 / 13.44
0.6107 / 0.6066
148.76 / 116.91
13.13 / 13.27
0.5305 / 0.5497
91.41 / 86.21
Ours
2.0B
14.31 / 14.12
0.4952 / 0.5103
114.99 / 107.44
13.78 / 13.99
0.4468 / 0.4752
79.65 / 72.33
Table 1: Quantitative comparison on iPhone and iPhone-PTZ benchmark . Results for short clips are on the left, full videos are on the right. The best results are highlighted in bold .
Method
iPhone [ 9 ] / iPhone-PTZ
Subject Consistency ↑
Background Consistency ↑
Temporal Flickering ↑
Motion Smoothness ↑
Aesthetic Quality ↑
Imaging Quality ↑
Gen3C [ 24 ]
0.8510 / 0.8228
0.8900 / 0.8597
0.9627 / 0.9376
0.9804 / 0.9819
0.4194 / 0.4222
0.6703 / 0.7304
Ours
0.9400 / 0.8574
0.9489 / 0.8816
0.9801 / 0.9478
0.9874 / 0.9844
0.4298 / 0.4435
0.7645 / 0.7557
Table 2: VBench comparison on iPhone and iPhone-PTZ benchmark . Results for iPhone are on the left, iPhone-PTZ are on the right.. The best results are highlighted in bold .
Figure 7: Qualitative comparison of short clips on the iPhone and iPhone-PTZ dataset .
Figure 8: Qualitative comparison of full videos on iPhone-PTZ dataset .
PSNR ↑
LPIPS ↓
FID ↓
Ours
13.99
0.4752
72.33
w/o Plücker
13.18
0.4897
78.23
w/o Source
13.04
0.5134
92.90
w/o Hybrid Warping
12.18
0.5347
84.75
Table 3: Ablation on hybrid warping and CCDM conditions .
Figure 9: Effectiveness of hybrid warping scheme . We compare per-frame warping (a) and our hybrid warping (c), along with their corresponding final results (b) and (d). Hybrid warping yields more complete and source-aligned coarse frames, leading to higher-quality final generations.
PSNR ↑
Subject Consis. ↑
Background Consis. ↑
Ours
13.99
0.8574
0.8816
w/o History Guidance
13.39
0.8543
0.8780
w/o Progressive Update
12.86
0.8487
0.8777
Table 4: Ablation on history-guided autoregressive generation .
Figure 10: Effectiveness of components in the history-guided autoregressive generation . Starting from the same first frame (a), we compare the last-frame outputs of our full model (b), the variant without progressive world-cache update (c), and the variant without history guidance (d).
Figure 11: More Qualitative comparison on the iPhone and iPhone-PTZ dataset .
Figure 12: Qualitative results on in-the-wild long video generation .