What did not exist two years ago
Video generation was a one-button operation for a long time: prompt in, clip out, and if it is no good, start over. That is expensive, because with video the fault is almost always partial — the scene is fine, there is just one object that does not belong, or the character's hand goes wrong for two seconds.
By 2026 that has changed. There is now a separate video editing capability, and a separate independent leaderboard for it, which in itself says this is not the same task as generating. Interestingly, the two lists do not line up: the model that leads at generating is not necessarily ahead at editing — and the other way round.
The four operations available today
- Removing an object. You take something out of the moving image, and the model fills in the background behind it, right through the clip.
- Replacing an element. The same thing, except it does not leave a gap, it puts something else there.
- Style or environment swap. The scene stays the same, but day turns to night, or the season changes.
- Extension. Continuing the clip. This is the weakest link — more on that in a moment.
The mask that is not there
Here comes the surprise if you are arriving from image editing. At the leading tools, video repair does not work by masking. You do not draw around anything: you describe in words what you want taken out.
One provider's documentation puts it exactly like this: you describe what is to be removed, and the model identifies it, tracks it through the clip and rebuilds the frame. The tracking is the point — that is what an editor would do by hand, frame by frame, and it is the work that used to eat days. The limits are modest: thirty seconds and 1080p is the typical ceiling for one operation.
Where the EU user hits a wall
This is worth knowing at planning time, not discovering mid-job. At one large provider, editing and extending an uploaded video is not available in the European Economic Area, Switzerland and the United Kingdom. Generation works; editing uploaded material does not.
This is not a technical limit but regulatory caution — and it means that if you work from the EU, this belongs among your tool-selection criteria. At the same provider, extension has a documented side effect too: the continuation rewrites the last frames of the input video. So it does not only add, it also touches what was already fine.
Several shots from one request
The other new thing is multi-shot generation: you do not ask for one clip but for a sequence, broken into shots. At one provider this is a field-by-field setup — duration, camera position, perspective and camera movement per shot — which is essentially a screenplay form.
It is convenient, but do not expect an editor's work from it. Sameness between shots drifts just as much as with separate generations, and the rhythm will not come out right on its own either. What you save is the handling of many separate requests — the planning and the selection stay yours. The whole workflow is what Building a One-Minute Video covers, in a separate post.
What to fix and what to regenerate
The decision comes down to a simple rule: fix it if the fault can be circumscribed; regenerate if it cannot.
- Fix: an object that does not belong, a caption, a logo, a bin in the background. These can be described, and the model finds them.
- Fix: the light or the season, if the scene is otherwise good. That is a global but uniform change.
- Regenerate: if the composition is wrong, if the camera move is not the one you asked for, or if the character is off. Those cannot be patched.
- Regenerate: if it is still not right after three rounds of fixes. The drift measured in image editing operates here too, only more expensively.
Common misconceptions
- "Video inpainting is the same as the image kind." It is not: there you draw a mask, here you describe. And here the hard part is the tracking, not the filling.
- "Editing is cheaper than regenerating." In time yes, in money not always — editing is billed by the second as well.
- "If this is the best model at generating, it is the best at editing too." The two leaderboards do not line up. Worth looking at separately; that is what Choosing an AI Video Tool sets out, tool by tool.
- "Extension does not touch what is already there." It does: one vendor writes itself that the last frames of the input video change.
What to take away
Regional repair is a real, working capability in 2026, and it removes exactly the pain that made AI video expensive. Keep two things in mind: the repair runs on words, not masks, and from the EU, editing uploaded material is not available everywhere. If the fault cannot be circumscribed, do not patch — regenerate.