Adobe and KAIST unveil MoviEdit, an AI tool for multi-shot video edits
Adobe Research and KAIST describe MoviEdit, a two-stage AI video editor that keeps edits consistent across scene cuts in clips up to 25 seconds long.
· 2 min read · By Efe Anidi
On this page +
Researchers from Adobe and South Korea’s KAIST published MoviEdit on 2 November 2026, a research framework built to edit long, multi-shot video — clips with several scene cuts and camera changes — rather than the single continuous takes most AI video editors handle. An edit such as swapping a character’s clothing or recolouring a background has to hold steady as the footage cuts between angles; the authors report that existing text-to-video editors tend to lose that consistency across cuts.
How it edits across cuts
MoviEdit uses a two-stage cascaded design built on a diffusion transformer with a mixture-of-experts backbone. A lightweight “draft editor” first runs bidirectional diffusion over a spatiotemporally subsampled version of the clip to settle on a globally consistent edit. A “render editor” then works through the footage causally, in chunks, refining that draft to full resolution while staying anchored to the source video and recent context. The two stages share weights and differ mainly in attention pattern and resolution.
Edits can be specified as text (“say it”), by editing one or more reference frames directly (“show it”), or both, and the same instruction can be applied consistently across a batch of clips. Output renders first as a 270p, 12fps draft, then a 540p, 24fps final cut, for clips up to 25 seconds long, whether single-take or multi-shot.
Training data and a new benchmark
Training used 86,619 filtered paired editing videos. They come from 61,251 raw clips of three kinds (multi-camera recordings, cinematic multi-shot footage and generated multi-shot content), which three automated pipelines turned into before-and-after pairs without manual annotation.
The authors also built MoviEdit-Bench: 64 generated multi-shot source videos across eight categories, with 128 editing scenarios, three-quarters of them local edits (adding, removing or swapping an object, or changing an appearance) and the rest global (changing environment or style). Against eight other systems, among them Aleph 2.0, HappyHorse-1.0, Gemini Omni Flash, Wan-Edit and Lucy Edit, MoviEdit scored the highest overall edit-success rate on text-guided edits, 0.790 versus 0.764 for the closest rival, Aleph 2.0; on image-guided edits the two models were close to even. Separately, the paper’s own inference optimisations (sliding-window attention, KV caching, few-step distillation) cut denoising latency for a 540p, 15-second clip from 364.1 to 11.8 seconds — a 30.9x speedup — on eight H100 GPUs.
Why it matters for transmedia and story tools
Multi-shot consistency is exactly the obstacle facing anyone using generative video for actual storytelling rather than single clips: a trailer, a web series, a sequence in a branching narrative. A model that can take a rough edit instruction and hold it across cuts, without re-rendering a whole sequence by hand, points towards editing tools that treat a video project more like a document you revise than a single clip you regenerate from scratch.
What’s still unconfirmed
MoviEdit is a research paper, not a shipped product. Adobe and KAIST have not released code or model weights, and the paper says nothing about releasing either. What you can get today is the benchmark: MoviEdit-Bench is on GitHub, linked from the project page, with the test videos, instructions and scoring code, so other teams can run their own editors against it. The comparison results are the authors’ own benchmark figures, not an independent evaluation, and full training and architectural details beyond what the paper discloses are unknown.
Questions and answers
What is MoviEdit?
MoviEdit is a research AI framework from Adobe and KAIST for editing long, multi-shot videos — up to 25 seconds, across several scene cuts — so that an edit stays consistent as the camera cuts between shots, rather than drifting the way single-shot video editors tend to.
Can I use MoviEdit today, or is it open source?
Not the model. The authors have not released MoviEdit's code or weights. They have published the MoviEdit-Bench evaluation benchmark on GitHub (adobe-research/MoviEdit-Bench), with download links for the source videos, edit instructions and evaluation code, plus MoviEdit's own reference outputs.
How does MoviEdit compare with other video editors?
On the authors' own MoviEdit-Bench, MoviEdit scored 0.790 overall edit success on text-guided edits, ahead of eight other systems including Aleph 2.0 (0.764), HappyHorse-1.0, Gemini Omni Flash, Wan-Edit and Lucy Edit. On image-guided edits it was roughly level with Aleph 2.0. These are the authors' own benchmark results, not an independent evaluation.
Sources:
