A change that looks convincing in one shot can fail as soon as the camera cuts: the subject, viewpoint, and lighting may all shift. The harder problem is not making an edit once, but carrying its intent through those changes without losing coherence. This work tackles that problem at both the data and modeling levels.
Key Findings
- Three automated data pipelines produce 87K paired editing videos from synchronized multi-view footage, real multi-shot videos, and generated content. This avoids relying on manual annotation to scale cross-shot supervision.
- A bidirectional stage first drafts a globally coherent edit on a compressed video representation; an autoregressive stage then propagates and refines it across the full video. The split balances whole-clip planning with detailed rendering.
- The authors introduce a benchmark for evaluating long-form multi-shot edits and report results above existing open-source models and on par with industry frontier models across text- and image-guided tasks.
Method and Evaluation
The central design choice is to separate global consistency from full-resolution refinement rather than ask one pass to solve both. The benchmark provides a dedicated way to measure whether an edit remains consistent across shots, alongside other editing qualities. Its quantitative source videos are generated, however, so benchmark performance alone does not establish robustness on the full variety of real-world footage.
Who Should Care
A good fit for researchers working on video editing, temporal consistency, or evaluation datasets, especially those studying edits that must survive scene cuts. Look elsewhere if you need a production-ready editor or evidence of performance on diverse real footage: the results are research claims, and the benchmark’s generated sources limit how far its scores can be generalized.