MoviEdit: Long-form Multi-Shot Video Editing Made Easy
Organizations: Adobe · KAIST
Abstract
Long-form multi-shot video editing requires an intended edit to remain coherent across abrupt changes in scene, viewpoint, and subject appearance, while processing substantially longer visual context than conventional single-shot editors. We introduce MoviEdit, an end-to-end framework for scalable long-form multi-shot video editing that jointly addresses the challenges of data, modeling, and evaluation. To provide cross-shot supervision data at scale, we develop automatic data-generation pipelines spanning synchronized multi-view footage, real-world multi-shot videos, and generated multi-shot content, yielding 87K paired editing videos without manual annotation. Architecturally, MoviEdit employs a cascaded bidirectional-autoregressive design: a lightweight bidirectional stage first establishes a globally coherent editing draft over a spatiotemporally compressed representation, which is then autoregressively propagated and refined over the full video. We further introduce MoviEdit-Bench, a dedicated benchmark for systematic evaluation of long-form multi-shot video editing. Extensive experiments show that MoviEdit substantially outperforms existing open-source models and performs on par with industry frontier models across text-guided and image-guided editing. Visit our project page for complete results and to access the benchmark: https://multi-shot-edit.github.io/
Figures & tables
| 17 | Model | Non-subject Preservation | Edit Localization | Subject Preservation | X-shot Edit Consistency | Edit Prompt Alignment | Video Quality | Overall Edit Success | ||||||||
| LPIPS | Precision | Recall | MFS | LPIPS | VLM | VLM | CLIP | ViCLIP | VLM | DINO | NIQE | PickS. | VLM | VLM | ||
| Open source | SDEdit (HoloCine) | 0.311 | 0.321 | 0.409 | 0.535 | 0.364 | 0.315 | 0.894 | 0.222 | 0.184 | 0.491 | 0.988 | 5.56 | 19.4 | 0.711 | 0.193 |
| Memory-V2V | 0.179 | 0.525 | 0.252 | 0.630 | 0.199 | 0.625 | 0.452 | 0.215 | 0.174 | 0.437 | 0.985 | 4.82 | 19.2 | 0.289 | 0.215 | |
| Lucy Edit | 0.174 | 0.729 | 0.192 | 0.647 | 0.185 | 0.751 | 0.750 | 0.213 | 0.171 | 0.351 | 0.984 | 5.82 | 19.2 | 0.508 | 0.235 | |
| SANA-Streaming | 0.141 | 0.651 | 0.430 | 0.651 | 0.212 | 0.707 | 0.490 | 0.219 | 0.181 | 0.592 | 0.984 | 5.48 | 19.3 | 0.504 | 0.354 | |
| Wan-Edit | 0.0816 | 0.840 | 0.355 | 0.686 | 0.103 | 0.809 | 0.692 | 0.217 | 0.178 | 0.515 | 0.983 | 5.38 | 19.3 | 0.568 | 0.378 | |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Pipeline | Data source | Editing task | # Pairs |
| Localization | Multi-camera | Highlighting | 3,500 |
| Outlining | 3,500 | ||
| Color filling | 3,500 | ||
| Subtotal | 10,500 | ||
| Pixel-aligned | Multi-camera | Add | 2,162 |
| Remove | 3,209 |
| Category | Composition | # Videos | Source-video design | Primary evaluation focus |
| C1 | Single human | 6 | One persistent subject observed across varying actions, poses, framings, and viewpoints. | Cross-shot localization under pose and viewpoint changes. |
| C2 | Two isolated humans | 6 | Two humans belong to the same scene, while each shot contains only one of them. | Target selection across shots and correct exclusion when the target is absent. |
| C3A | Two co-visible humans, same gender | 6 | Same-gender subjects are distinguished by roles, clothing, hairstyles, accessories, positions, or object ownership. | Subject disambiguation between visually and semantically competing humans. |
| C3B | Two co-visible humans, different gender | 6 | Different-gender subjects co-occur within shots and are distinguished by roles and visual attributes. | Cross-shot subject binding with multiple co-visible candidates. |
| C4 | Three humans | 6 | Three text-addressable characters share a continuous scene with distinct visual or semantic identifiers. | Localization under increased multi-person competition. |
| C5 | Four-or-more-person ensemble | 6 | Larger ensembles contain distinguishable roles, appearance, props, and spatial positions. | Target localization in crowded multi-person scenes. |
| Edit task | Scope | # Scenarios | Ratio |
| Add | Local | 24 | 18.75% |
| Remove | Local | 24 | 18.75% |
| Appearance change | Local | 24 | 18.75% |
| Swap | Local | 24 | 18.75% |
| Environment change | Global | 24 | 18.75% |
| Stylization | Global | 8 | 6.25% |
| Local | Global | ||||||
| Dimension | Metric | Add | Remove | App. | Swap | Env. | Style |
| Non-subject preservation | LPIPS | ✓ | ✓ | ✓ | ✓ | – | – |
| Edit localization | Precision (VLM) | ✓ | ✓ | ✓ | ✓ | – | – |
| Recall (VLM) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |
| Subject preservation | MFS | ✓ | – | ✓ | ✓ | ✓ | ✓ |
| LPIPS | ✓ | – | ✓ | – | – | – | |
| FiVE-Bench | VPBench | |||||||||||||||||
| Model | Structure | Background Preservation | Text Alignment | IQA | Motion | FiVE-Acc | Background Preservation | Text Alignment | Quality | |||||||||
| Dist. | PSNR | LPIPS | MSE | SSIM | CLIPS. | CLIPS. edit | NIQE | Fidelity | VLM | PSNR | SSIM | LPIPS | MSE | CLIPS. | CLIPS. edit | FVID | ||
| Open | Lucy Edit | 15.49 | 25.66 | 99.65 | 50.52 | 82.15 | 25.41 | 20.33 | 6.30 | 87.95 | 0.3532 | 26.53 | 0.856 | 0.091 | 0.0095 | 11.11 | 13.89 | 0.236 |
| SANA-Streaming | 14.63 | 24.25 | 111.66 | 56.80 | 80.52 | 24.82 | 19.78 | 6.22 | 86.44 | 0.1921 | 23.34 | 0.825 | 0.108 | 0.0100 | 11.96 | 13.86 | 0.252 | |
| Wan-Edit | 11.36 | 25.38 | 96.88 | 34.92 | 81.71 | 26.64 | 21.18 | 6.16 | 88.20 | 0.4010 | 27.09 | 0.877 | 0.076 | 0.0023 | 11.00 | 13.54 | 0.238 | |
| Private | Gemini Omni Flash | 31.29 | 20.45 | 132.49 | 129.43 | 67.96 | 28.02 | 22.24 | 6.09 | 82.27 | 0.8007 | 23.44 | 0.789 | 0.102 | 0.0114 | 11.46 | 13.30 | 0.270 |
Video Editing
Video editing uses AI to alter, compose, relight, or reorganize moving footage while controlling how changes persist across frames.
See all 188 papers