Genjutsu AI video: change what is in a clip without reshooting it

The reason Genjutsu output looks like your footage rather than a re-imagining of it is the mode it runs in. Seedance 2.5 can use a reference video two ways. In generation mode the clip is a loose guide: the model keeps the story beats but re-invents framing, timing and faces. In editing mode the clip is the ground truth — the output inherits its exact length and aspect ratio, and only what the instruction names is allowed to change. Genjutsu uses editing mode exclusively. We tested the other path and it produced a good video of the wrong thing: containers and cranes in the right order, but every frame reframed and every player recast.

Concretely, four things are preserved on every job: identity (faces, bodies, the clothes you did not name), performance (every movement at its original speed), camera (the path, the framing, the cuts) and timing (a 15.07-second clip comes back at 15.04 seconds). What changes is exactly the set of things you labelled — or, in Motion Transfer, the environment around the people. This is why the two presets are worth keeping separate in your head: they are the same engine with opposite instructions about what "only" refers to.

The model is obedient to the instruction it receives, so the instruction is where the craft is. VdoBloom builds it from your labels: "@Image1 replaces the striker's kit; @Image2 replaces the advertising boards behind the goal — match each reference photo's look, material and colour — do not change the clothing or appearance of anyone not listed." You never see or write that; you see one label field per photo. A vague label ("the player") gives the model latitude it will use; a specific one ("the striker in the number 10 shirt") does not.

Single-target swaps in shots with a clear subject are the reliable case: an outfit on the person in frame, a product in a hand, a prop on a table. Our reference test — a kit swapped for a long leather coat across a dribble, a shot, a knee-slide and a close-up — held the face, the stride and the coat's belt across all five sampled frames, and the frames where the target was out of shot came back near-identical to the source.

Motion Transfer is strongest when the new location is described by two or three wide photos and a short line of text. In the stadium-to-port test the model kept all eleven players, the referee, the ball and the boom microphone while rebuilding the ground, the stands and the skyline; the more the photos agree about the place, the more coherent the result across the clip.

The weak case is crowds of similar people. Asked to dress two named players in the same coat, the model did — and also touched an un-named player in the wide shot. The mitigations are the ones the interface pushes you toward: one distinct photo per target, a specific label for each, and the note field for anything that must not change. It is also why a 480p draft is the right first run on a busy shot: a 15-second draft is 212 credits, and you re-run only the keeper at 720p or 1080p.

Genjutsu is billed per second of clip on Seedance 2.5's with-reference-video rate, at the output resolution you choose. Because the output is exactly as long as the input, the price is a function of two numbers you know before you start: clip length and resolution. A 15-second clip is 212 credits at 480p, 473 at 720p or 852 at 1080p; the 30-second maximum at 720p is 912. The tab reads the clip's length from the file the moment you pick it and shows the figure before the upload finishes.

Seedance's editing mode has hard input limits that it enforces only after a job is submitted: 2-30 seconds, 480p-720p sources, 24-60 frames per second, under 200 MB. VdoBloom probes the clip and refuses anything outside those limits before credits are deducted, and it reads the real average frame rate rather than the nominal one, so variable-frame-rate phone recordings that report 600 fps are not wrongly rejected. If a job fails at the model after passing those checks, the credits are refunded automatically and the reason the model gave is shown to you rather than a generic error.

There is no plan gate. Genjutsu is available on free accounts (though the 10 free signup credits do not cover a job), on subscriptions and on one-time packs from $2.49 whose credits never expire. That is the main practical difference from Higgsfield, where Genjutsu needs Seedance 2.5 and therefore at least the Pro plan; the head-to-head is at /tools/higgsfield-genjutsu-alternative/.