AI video editing: prompts, video to video, object and background removal

Which tool for which job

Editing video with a prompt

A prompt edit changes part of a clip while the motion stays. Write it as an order: what changes, and what must not. “Replace the two trucks with two elephants in the same place, and keep the other shots unchanged.” Always ask yourself afterwards whether it really edited, or quietly re-generated the shot. Play it next to the source: if people stand still or the framing changes, it was a new generation.

Edits bill input and output, so trim the source to the shot you are changing. On Kling, an input longer than the output adds to the bill.

Video to video

Our recast test ran a 16-second commercial with eleven cuts through four routes with the same approved pictures. Kling 4.0 Edit and Wan 3.0 kept the cuts and gestures, Higgsfield Genjutsu Lite handled the whole 16 seconds in one pass, and Seedance 2.5 Preview in Dreamina’s multimodal mode re-staged the shot. To change the format as well, edit from a horizontal original rather than outpainting a vertical one.

Removing a video background

Two approaches exist, and they suit different needs. Matting finds the person and keeps their original pixels, so the face is unchanged, and you place them over anything. Generative replacement redraws the room and can match the light, but it redraws the person as well.

All three struggled with fine hair against a bright wall. Keep overlays out of the clip you send, and add them back in the edit. So for a free video background remover, the honest answer is that open models work and cost time and edge quality. The full test shows the clips.

Removing an object from a video

Our object-editing note covers six tools. For removal, name the object in a short prompt and say everything else stays unchanged. Seedance 2.5 Video Edit takes a text prompt, and some hosts add a way to mark the object on a frame where it is fully visible. We have not run a removal with it.

On our own job with Genjutsu, the model missed about half the objects we asked it to change, so check every frame.

Free online object removers exist, and we have not tested them. Expect limits on length and resolution, and a watermark. We will add a test before recommending one.

Face swap and replacing a person

Our experience here is recasting whole people across a commercial, with new faces approved as pictures first. We have not tested single-purpose face-swap apps. A recast keeps the performance and changes who is performing, so it raises the same questions as any face swap: use only people who have agreed, and do not depict a real person doing something they did not do. Some platforms also refuse clips of real people, as fal’s Seedance 2.5 did in our test.

Style transfer

Restyling means keeping the motion and changing the whole look. Our nearest evidence is the room-and-outfit restyle, where Seedance 2.5 Video Edit took one reference still as the target look. Google says Gemini Omni Flash can edit uploaded videos with new prompts and ingredients, at about $0.10 a second of 720p output on its API, and we have not run it for this. We have not run a controlled test of artistic style transfer, such as live action into anime, so we make no pick for it yet.

Do you need an AI video editor at all?

For cuts, captions and sound, a normal editor is still quicker. Use AI when you are changing what is in the frame. CapCut and Canva now build generation into the editor, as our roundup covers, and our platform work aims at the same gap for production use. For polish, read the upscaling note before you reach for 4K.

Written by Oleksiy Onishchenko, who produces AI films for brands and teaches AI video workshops. Founder of easyai.video.