Paper review · November 5, 2026
MoviEdit: Making Long-Form Multi-Shot Video Editing Easy
MoviEdit: Long-form Multi-Shot Video Editing Made Easy
The authors built 87K pairs of automatically generated multi-shot editing data, a cascaded architecture that drafts bidirectionally and then renders autoregressively, and a dedicated benchmark, enabling long-form video editing that stays consistent across cuts.
Problem
Existing video editing models are designed for single shots a few seconds long. They struggle to keep edits consistent across cuts where the scene, viewpoint and subject appearance change. Paired training data, a model architecture that can handle long context, and an evaluation benchmark for this were all missing. Processing the entire source and output videos together makes attention cost unmanageably large.
Method
On the data side, the authors built three pipelines that create paired data without human labels: (1) pixel-aligned editing, where the first frame of each shot is tiled into a grid, edited at once with an image editor, and then propagated within each shot by a single-shot video editor; (2) subject localization based on segmentation masks; and (3) quasi-pixel-aligned editing that shares noise and trajectories in a multi-shot generator and uses ControlNet. A VLM quality filter was applied at each stage. The model consists of two stages that share the same weights: a Draft Editor that produces a global edit draft with bidirectional diffusion on a spatiotemporally subsampled representation, and a Render Editor that performs chunk-wise causal diffusion at full resolution conditioned on the draft. The authors also present MoviEdit-Bench.
Results
- The authors created 87K pairs of multi-shot edited video data without manual annotation. In the filtering stages, 17%/38% of the multi-camera pairs and 32%/47% of the movie clip pairs were filtered out.
- The authors report that the model substantially outperforms existing open-source models in both text-based and image-based editing and is on par with commercial frontier systems.
- In a small stress test, autoregressive editing without draft conditioning accumulated appearance drift and oversaturation, while draft-conditioned rollouts kept the edited appearance stable.
- The authors show examples of generalization to videos longer than those seen in training.
Limitations
In the provided text, the quantitative results on MoviEdit-Bench are cut off, so the size of the claim of being "on par" with commercial models cannot be checked. Part of the data relies on in-house synchronized multi-camera footage. Evaluation was done mainly on the authors' own benchmark. Text-based and image-based editing were trained as separate models.
Why it matters
This work presents the data, model and evaluation needed to extend video editing AI from short clips to real video productions made of multiple cuts, all together. It is a useful reference for teams looking to add generative editing to advertising, film and content production tools.
Key terms
- Multi-shot video editing
- The task of making the same edit apply consistently to the same subject across a video made of multiple cuts, even when the viewpoint or scene changes.
- Causal Diffusion Forcing
- A diffusion approach that splits a video into chunks, assigns a different noise level to each chunk, and trains the model to generate the next chunk sequentially by looking only at previous chunks.
- Cascaded bidirectional-autoregressive architecture
- A design that first creates a global draft bidirectionally over the whole video at low resolution, then uses it as a reference to generate the high-resolution video sequentially chunk by chunk, addressing both memory and consistency issues.
Original
Authors: Suhyeon Lee, Zhengang Li, Soo Ye Kim, Peihao Zhu, Jong Chul Ye, Chun-Hao Paul Huang, Duygu Ceylan, Yan Kang, et al. · Posted 2026-11-02 · License CC BY-NC-ND 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.