Tell MoviEdit to "remove the man" from a 15-second clip with multiple cuts, and he vanishes from every single shot, across every camera angle. Built by Adobe and KAIST, this video editing model keeps a single edit consistent across an entire multi-shot clip. For now, though, the GitHub repo only releases the benchmark for this task—not the model itself.
The man in the black suit removed from every shot; the small inset at the top left is the original clip.
What's the actual problem?
Current AI video editors are built for short, single-shot clips. In a real film, the camera constantly cuts, changing angles and lighting on the fly. After every cut, the editor has to find the exact same person again and keep their appearance identical to the previous shot. If you edit each shot separately, that consistency falls apart. On the flip side, if you try to process the whole video in one go, memory usage and compute costs blow up.
A clever two-step fix
Here is how they solved it: first, the model edits a lightweight, low-res version of the entire video (using a "Draft Editor" at 270p and 12 fps). Because this step sees every shot at once, it locks in a consistent global plan for the edit. Next, the "Render Editor" generates the video chunk by chunk at full quality (540p, 24 fps), using that rough draft as an anchor so the look doesn't drift over time. Both stages actually share the same model weights—at least according to the authors' own breakdown.
The original video alongside three different edits.
The speed is pretty impressive too: after distillation and a few optimizations, editing a 15-second clip on eight H100 GPUs takes about 12 seconds, down from 364 seconds without them. They also built the training dataset without any manual labeling—87K before-and-after video pairs automatically filtered to prevent visual quality from degrading over longer sequences.
What can it actually do?
Edit using text prompts, or tweak a single frame from any shot and let the model propagate that change across the whole video (it accepts multiple reference frames too).
Handle 25-second videos smoothly, even though it was only trained on 15-second clips.
Apply a single consistent edit across several completely separate videos at once.
One video, three different edits: a character swap, grass growing on the subway floor, and a darker mood.
What's inside the GitHub repo?
The repository only includes MoviEdit-Bench: 64 AI-generated multi-shot videos (4 to 6 shots each, around 15 seconds long) paired with 128 editing scenarios, covering everything from adding and removing objects to changing environments and visual styles. You get the data, evaluation scripts, and reference results. Models are scored using seven metrics (checking things like scene preservation and cross-shot consistency) alongside an AI judge. The code is licensed for noncommercial research only, and the inference code to run the model itself isn't included.
What the paper reports
In their self-reported numbers, MoviEdit scores 0.790 in overall edit success for text-guided editing, edging out Runway Aleph 2.0 at 0.764, while the best open-source model sits at 0.378. For frame-guided editing, Aleph 2.0 is slightly ahead when using the first frame (0.806 vs. 0.803), whereas MoviEdit takes a slight lead when you pick an arbitrary reference frame (0.786 vs. 0.781). The main catch they point out: when there are multiple look-alike people in the frame, or the target is only partially visible, the edit can accidentally jump onto the wrong person.
My take
The best part here is the first approach: make a fast, lightweight global decision across all shots first, then fill in the high-res details. That's a pattern that could easily transfer to other long-form generation problems. The caveat is that every benchmark comparison runs on a test set the authors built themselves, relying on AI-generated videos and an AI judge. Until the actual model weights are released, nobody can verify these numbers independently. For now, the real value of the GitHub repo is giving other researchers a standard yardstick to test their own multi-shot editors.
