The first thing to say is that the button this trend is usually described around is not on the web version. The tools page lists AI Inpainting and Face Cutout under images; there is no face swap among them. What there is, and what did the work here, is something the description does not mention: the composer takes two reference images at once.
Two references, one sentence
Two pictures went in separately on the AI image generator at 16:9 — a man in a suit with his arms wide in a dark stadium, lit cool and blue, and a plain portrait of a different man lit warm and flat. Then both were attached to one request.
The swap. Use the second reference image as the face. Put that young man's face and head onto the man in the first reference image, so it is now him standing in the suit with his arms spread wide in the dark stadium. Keep the pose, the suit, the stadium and the cool blue lighting from the first image exactly as they are.
That worked on the first attempt. The pose, the suit and the stadium came through unchanged and the face is the one from the portrait.
Two of the warnings, checked
The cheek comes from the portrait; the hands are the template's. The two should match if the swap matched the light.
The skin tone missed. In the original picture the face and the hands are the same person under the same light, and they sit 3.86 apart in color. After the swap they sit 14.49 apart — close to four times as far. The number that moved is warmth: the face arrived at b* of +1.2 where the template's face had been −7.3, while the hands stayed where they were at −4.3. The face brought its own warm indoor lighting with it and the blue stadium did not overrule it.
The jaw seam did not happen. Measuring the strongest row of edge along the jaw line gives 28.54 before the swap and 22.58 after — slightly softer, not sharper. For scale, planting a deliberate three-pixel seam of twenty levels into that same band pushes the number to 99.8, so a seam of the kind people worry about would be unmissable on this measure. There is nothing there.
So of the two things you are told to fix by hand after a swap, one is real and one was not a problem in this run. The color is the one worth the work, and it is a color adjustment on the face layer rather than anything clever.
Then it moves
The swapped still went to the AI video generator as the first frame of a five-second clip. The prompt asked for the pose to be held and said in plain words that the hands and fingers do not move.
Five seconds. The camera holds, the light pulses, and the hands are supposed to be still.
Each hand is matched against its own shape in the first frame. Both fall away within the first two seconds and stay down.
They do not stay still. Both hands fall from a perfect match to 0.87 to 0.90 within about two seconds and sit there for the rest of the clip, finishing at 0.894 and 0.907. To know whether that is a lot, the same measure was pointed at the reference hand deliberately warped until it looks twisted: that reads 0.953. The drift in the clip is further from the original than a deformation you would have no trouble spotting by eye.
What changes is not position — the hands stay where they were put. It is the fingers and the palm, which resolve differently as the light comes up on them. That is why telling it not to move the hands does not help: nothing is moving. The shape is being redecided.
What to do with that
Where this stops
One swap, one clip, one model. Each figure is a single run, so the sizes are indicative — the fourfold color miss is large enough to plan around, and the missing jaw seam is one negative result rather than a guarantee that seams do not happen.
The hand measurement has a narrower reading than it might look. It says the shape inside a fixed box stops matching. A viewer watching at speed may not care, and the clip above is not obviously broken. The number is useful because it tells you when it happens — about two seconds in — which is something you can edit around.
Three warnings, one worth your time
This edit comes with three pieces of received advice: fix the skin tone, patch the jaw, expect the hands to go. Measured, they sort out unevenly. The jaw did not need touching. The skin tone needed exactly the one adjustment it is usually given, and needed it badly. The hands went further than the warning suggests and could not be prompted out of it, which moves them from something to fix into something to shoot around. If you have an hour for this edit, none of it belongs at the jaw, ten minutes belong on the face color, and the rest belongs in deciding how long the hands are on screen.
Configuration and cost. Three stills: Image, Seedream 4.5, 16:9, 2K, Standard — the panel declared 1 credit per send. The clip: Video, Seedance 2.0 Mini, 16:9, 5s, 720p, Standard — the panel declared 25 credits. Total 28 credits, read from the send panel on each send, 2026-09-21. Credit costs are tied to that configuration on that date.
Files.tpl.png (the pose), src.png (the face), swap.png (the two combined), all 2560×1440; clip.mp4 1280×720, 24 fps, 121 frames, 5.04 s, shown on this page as clip.gif.
The prompt block is reproduced as it was sent.
How the numbers were taken. Colour distance is the Euclidean distance between median CIELAB values of a cheek patch and the two palms, all drawn in the figure and checked before use. The jaw measure takes the vertical edge response across a band at the jaw line and reports the strongest single row, because a seam is one row that is steeper than its neighbours; a median over the band was tried first and discarded, since planting seams of 8, 20 and 40 levels moved it by less than two either way. The row measure was then self-tested with the same planted seams and returned 63, 100 and 180 against a 63 baseline. The hand measure is multi-scale normalized cross-correlation of each hand against its own shape in the first frame, self-tested against known rescalings, which it recovered exactly, and against deliberate warping of the reference at three strengths, which dropped it to 0.996, 0.982 and 0.953. Every sampling box was drawn back onto its frame and checked; an earlier pair of hand boxes sat mostly on black background and was replaced.