Replace people and places in existing footage: which AI tools keep the motion

The test footage is a 16-second commercial with eleven cuts. The job was to keep every cut, camera move and gesture and replace everything else: five main people, five locations, the brand colour, and every subtitle, logo and floating graphic. Each new person and place was made first as a still and approved, and every tool below got the same stills.

The result was needed in 9:16 and in 16:9, which splits the work into two different jobs. Replacement changes what is in the frame and must keep the motion. Reframing turns a finished vertical cut into a wide one by inventing the missing sides. This note compares both on the same footage.

Replacement: what kept the motion

The test is simple. Play the result next to the source: does every person still move, turn and cross the frame at the same moment, and did the cuts stay where they were?

Seedance 2.5 kept a talking head best in our room and outfit test, so its failure here needs a word. In Dreamina, a video plus reference pictures goes through the multimodal mode, which generates a new clip guided by the video. There is no edit switch. The people and places were right; the performance was not.

Two routes were not run. H3 Max Recast on fal swaps people only, up to four, and would have left the locations, the colour and the text. Kling O3 Edit builds each replacement around one main character, so a shot with several people would need one run per person.

Wan 3.0 needs short runs

Wan 3.0 accepts up to 15 seconds of source video and ten reference pictures, so the obvious move is one run for the whole remainder of the clip. We tried it: six shots, nine pictures and a long list of instructions, 9.55 seconds in. Wan recast every person, but skipped most of the rest. Subtitles, a floating app card and the source shop’s signs and stickers survived, and two locations kept their original look.

Split into runs of two shots, each with only the two or three pictures those shots need, and with every leftover named in the prompt, the same shots came out clean.

Reframing: stretching the vertical cut to 16:9

The cheaper idea was to keep the approved vertical cut and outpaint it: put the 9:16 frame in the middle of a 16:9 canvas and let a model paint the left and right. The catch is geometry. A 9:16 frame fills only about a third of a 16:9 frame, so two thirds of every picture is invented.

The same shot, five ways

Three reframes of the vertical shot, then the same shot from the two edit routes, which started from the original horizontal footage. All five play at their native resolution.

The same middle frame of each result at full size: the left side, the man at the counter, the right side.

The edits win, and it is not close. An edit starts from a real wide frame, so its sides have real geometry, real depth and real people. A reframe has to invent them, and at two thirds of the frame the invention shows. Reframing is worth it for small additions, a strip of wall or sky, not for turning a vertical cut into a horizontal one.

Cost of every run

Every route is priced for 5 seconds of footage at standard list prices, so the rows compare directly. Higgsfield credits use the Plus plan billed monthly, $59 for 1,200 credits. Kling.ai credits use the Standard plan billed monthly, $8.80 for 660 credits. Rates change; check fal, Higgsfield and Kling.ai for current figures.

What to take away

1. Make the stills first. Every route above took the same approved people and places, which is what kept the cast consistent across tools and formats.
2. Check that the tool edits, not re-generates. Play the result beside the source. If people stand still or the framing changes, it was a new generation.
3. Keep each edit run small. Two or three shots, only the pictures those shots need, and every leftover named in the prompt. An overloaded run recasts the people and quietly skips the rest.
4. Pick per shot. Kling 4.0 for location swaps, Wan 3.0 when the background crowd matters. Both keep the cuts, so the shots mix cleanly.
5. Remake each format from its source. If a horizontal original exists, edit it. Outpainting a vertical cut invents two thirds of the frame.
6. Cut the unused shots before paying. Edits bill the input seconds too, so drop the shots the final cut does not use.

The prompt buttons show the structure of the prompts we used, with the shot-specific details replaced by placeholders in square brackets.

Remake this 16-second ad as if it were filmed in [new setting]. Keep every cut, camera move, framing, body movement, hand gesture and timing exactly as in the source; change only the people, clothing, locations, colours and text. Each replacement person keeps the face, hair and complete outfit of their image.

REFERENCES:
- @Image1 replaces [the person in shot A] ([time range]); the place becomes @Image6.
- @Image2 replaces [the person in shot B] ([time range]); the place becomes @Image7.
- ... one line per person and place, with the time ranges where they appear.

PEOPLE: everyone else is [what the extras should look like].
CLOTHING: [the wardrobe rule for extras].
COLOUR PALETTE: replace every [old brand colour] with [new palette]. No [old brand colour] anywhere.
PHONE SCREENS: [what every visible screen shows].
REMOVE COMPLETELY: all subtitles, small legal text, the source logo, floating graphics and all signs and lettering. Add no new text.

Edit [Video 1]: Keep every cut, camera move, framing, body movement, hand gesture and timing exactly as in the source; change only the people, clothing, locations, colours and text. Each replacement person keeps the face, hair and complete outfit of their image.

0.0-2.5s: [the person in shot A] becomes the person in [Image: person A] ([outfit]); the place becomes [Image: place A].
2.5-4.8s: [the person in shot B] becomes the person in [Image: person B]; the place becomes [Image: place B]. [The extra who crosses the frame] becomes [description].

Replace every [old brand colour] with [new palette]; no [old brand colour] anywhere. Remove all subtitles, small legal text, floating graphics and all signs and lettering; add no new text. Keep everything else unchanged. Real handheld footage, [light], no glossy CGI look.

Edit @Video1. Keep every cut, camera move, framing, body movement, hand gesture and timing exactly as in the source; change only the people, clothing, locations, colours and text. Each replacement person keeps the face, hair and complete outfit of their image.

First shot (0.0-2.5s): [the person in shot A] becomes the person in @Image1 ([outfit]); the place becomes [new place] in @Image2.
Second shot (2.5-4.8s): [the person in shot B] becomes the person in @Image3; the place becomes @Image4. [The extra who crosses the frame] becomes [description], with the same walk and timing.

Every other person is [what the extras should look like]. Replace every [old brand colour] with [new palette]; no [old brand colour] anywhere. Remove all subtitles, small legal text, floating graphics and all signs and lettering; add no new text. Keep everything else unchanged. Real handheld footage, [light], no glossy CGI look.

Edit Video 1: keep both shots, the cut, every camera move, hand and body movement and the timing exactly; change only what is listed.

Shot 1 (0.0-1.1s): replace [the person] with the person in Image 1, same face, hair and outfit. Change the [place] into the [new place] in Image 2: [three or four concrete features of the new place]. Remove every sign and sticker: [name each leftover from the source, word for word].
Shot 2 (1.1-2.2s): [the hands, props or screen and what they become]. [Any object that keeps the old colour] becomes [new colour].

Change every [old brand colour] to [new palette]; no [old brand colour] anywhere. Remove all subtitles including [the exact word on screen], floating chips and all lettering; add no new text. Real handheld footage, [light].

Continue the same [place] to the left and right: [materials, fittings and props already visible in the frame], and [what lies beyond it] with a few passers-by. Same camera, light and colour grade as the source. No text, signs or logos.