Genjutsu AI video: change what is in a clip without reshooting it
Genjutsu is VdoBloom's video-to-video editor. You upload a clip you already have, add one photo per thing you want to change, label each photo with what it replaces, and the output is the same shot with those changes applied. Every person you did not name, every movement, the camera path and the timing come back exactly as filmed. It runs on ByteDance Seedance 2.5's video-editing mode, the same model family behind Higgsfield's Genjutsu, reached at /dashboard/video-creation/genjutsu/.
It has two presets that are really two instructions to the same engine. Object Swap changes the things you label — a character, an outfit, a product, a prop — and leaves the scene alone. Motion Transfer does the inverse: it keeps every person and every move and rebuilds the location around them from your photos. In our own tests a footballer's kit became a leather coat with his face, stride and the ball untouched, and a stadium became a container port with all eleven players, their poses and even the boom microphone identical.
The practical rules are short. Clips must be 2 to 30 seconds and 480p to 720p; the output is as long as the clip and keeps its aspect ratio, so you pick only the output resolution — 480p, 720p or 1080p. You can attach up to ten reference photos, each with its own label, which is how you change two, three or ten things in one pass. Generation takes about four to seven minutes and is billed per second of clip on the with-reference-video rate.
What this tool does
- Genjutsu takes a 2-30 second source clip at 480p-720p and returns a clip of the same length and aspect ratio; you choose only the output resolution (480p, 720p or 1080p).
- Up to 10 reference photos per job, each carrying a label such as "the man on the left" or "the drink can on the table" — one labelled photo per thing you want to change.
- Two presets, one engine: Object Swap changes only what you label; Motion Transfer keeps every person and movement and replaces the location.
- Faces, poses, camera motion and timing are preserved — in VdoBloom's September 2026 tests a player's kit was swapped for a leather coat with his face and stride unchanged frame for frame.
- It runs on Seedance 2.5's video-editing task type (duration follows the source, ratio adaptive), the same ByteDance model that powers Higgsfield Genjutsu.
- Cost is per second of clip on the with-reference-video rate: a 15-second clip is 212 credits at 480p, 473 at 720p or 852 at 1080p; a 30-second clip at 720p is 912.
- The price is quoted from the clip's length the moment you pick the file — before the upload finishes and before anything is charged.
- Clips that Seedance would reject (over 30 s, 1080p+ sources, under 24 fps, over 200 MB) are refused before credits move, not after.
- Typical turnaround is four to seven minutes; the job keeps running if you close the tab and appears in My Creations.
- Credits from one-time packs never expire; there is no plan that unlocks or withholds Genjutsu.
What Genjutsu keeps and what it changes
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.
Where it works best, and where it slips
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.
Pricing, limits and the checks that run before you pay
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/.
How it works
1.Choose Object Swap or Motion Transfer
Object Swap when you want to change things inside the shot; Motion Transfer when you want the same people and action in a different place. Both cards on the tab show a real output from our tests.
2.Upload your clip
MP4 or MOV, 2 to 30 seconds, 480p to 720p (export a 720p version if your footage is 1080p or 4K). The length is read from the file immediately and the price appears at the bottom of the page.
3.Add one photo per change and label it
For Object Swap, a photo of the new outfit, product or character with a label saying what it replaces. For Motion Transfer, photos of the new location labelled by what they show. Up to ten; unlabelled photos are used as general reference only.
4.Add a line of guidance if you want
Optional. "Keep her hair as it is", "make the coat look worn", "a neon Tokyo street at night". A preset with labelled photos works without it.
5.Pick the output resolution and generate
480p for a draft, 720p for most social output, 1080p for delivery. The credit cost is shown before you confirm and deducted up front; if generation fails, it is refunded automatically.
Frequently asked questions
What is Genjutsu on VdoBloom?
A video-to-video editor at /dashboard/video-creation/genjutsu/. You give it a clip plus labelled reference photos and it returns the same clip with the labelled things changed — or, with Motion Transfer, the same people and moves in a new location. It runs on ByteDance Seedance 2.5's editing mode.
Is this the same as Higgsfield Genjutsu?
It is built on the same model family and does the same two jobs, Object Swap and Motion Transfer. The differences are in the workflow and the billing: VdoBloom lets you label each reference photo with what it replaces, accepts up to ten photos, quotes the price from the clip length before you upload, and sells one-time credits that never expire. A full comparison is at /tools/higgsfield-genjutsu-alternative/.
Can I change more than one person or object at once?
Yes — that is what the per-photo labels are for. Add one photo per target and write what each replaces. In our test with two named players both changed while their faces and movement stayed. In crowded shots the model can occasionally touch someone you did not name; distinct photos and specific labels ("the striker in the number 10 shirt") reduce that.
Does it keep the person's face?
Yes. Edit mode preserves identity, pose, camera motion and timing. Changing the face itself is possible with Object Swap by labelling a photo "the man's face", but the default behaviour is to keep it.
What clips does it accept?
MP4 or MOV, 2 to 30 seconds, 480p to 720p sources (854×480 to 1280×720), 24 to 60 fps, under 200 MB. Anything outside those limits is refused before you are charged. 1080p and 4K sources need to be exported at 720p first.
How much does Genjutsu cost?
It is billed per second of clip on the with-reference-video rate, at the output resolution you choose. A 15-second clip is 212 credits at 480p, 473 at 720p or 852 at 1080p; a 30-second clip at 720p is 912. The exact figure for your clip appears the moment you pick the file.
Can I try it on the free signup credits?
No. The cheapest possible job — a 4-second clip at 480p — is 41 credits, above the 10 credits a new account starts with. Credit packs start at $2.49 and never expire.
Object Swap or Motion Transfer — which do I need?
Ask what should stay. If the place stays and something in it changes, Object Swap. If the people and the action stay and the place changes, Motion Transfer. They are the same engine with a different instruction, so you can run both on the same clip.
Related tools
AI object swap: replace a person, outfit or product in a video and keep the shot
Replace a person, outfit or product in a video: upload the clip and a labelled photo of the replacement; the rest stays as filmed. 2-30s, up to 1080p.
AI motion transfer: keep the people and the moves, change the world around them
Keep every person and movement from your clip and rebuild the location from photos. Faces, poses and camera stay. 2-30s clips, up to 1080p.
Higgsfield Genjutsu alternative: the same engine with labelled targets and credits that don't expire
VdoBloom vs Higgsfield Genjutsu: same Seedance 2.5 engine. Differences: per-photo labels, price before upload, no plan gate, credits that never expire.
Seedance 2.5: the 30-second model, and how to run it
ByteDance Seedance 2.5 renders 4-30 seconds at 480p, 720p or 1080p (10-bit). Pick it in Text to Video, Image to Video or Effects. 30s at 480p is 276 credits.
Seedance 2.5 price: the full credit table, per second and per clip
Seedance 2.5 costs 9.2 credits/sec at 480p, 20 at 720p, 36 at 1080p. Cheapest render is 37 credits (4s, 480p); a full 30s at 1080p is 1,080.
Ready to try it?
Open Genjutsu