Higgsfield Genjutsu: recast a shot without reshooting it
Every other tool in this category starts from nothing and hopes the motion lands. Genjutsu starts from footage you already shot and keeps the part that is hardest to fake — how the camera moved, how the performance was timed. Higgsfield added it on 1 September 2026. This page covers what the two modes actually do, what a render really costs in credits, and the one line you should not cross.
These are the official sites for everything described on this page. Every link goes to the vendor's own page — we are not affiliated with any of them. We list them because a guide that tells you what something is but not where to get it has only done half the job.
What Genjutsu is
Genjutsu is a video-to-video tool inside Higgsfield, released on 1 September 2026 according to Higgsfield's own changelog entry. It is not a generator in the sense the rest of this site uses the word. You do not describe a scene and wait. You bring a clip, and Genjutsu rebuilds part of it — the person in the shot, the setting, or an object in the frame — while everything else stays as it was filmed.
That framing is the whole product. In the last two years the hard problem in AI video has quietly moved. It used to be fidelity: does the picture look real. It is now motion — getting a shot to move the way a director would have moved it, rather than the way a model happens to. Most tools solve this by generating from a text prompt and hoping the blocking lands on the first or third try. Genjutsu sidesteps it: film the motion once, properly, then reuse it.
Higgsfield calls it their strongest video transformation tool. The claim is consistent with what it does rather than with how it looks — nothing in Genjutsu's output is a new capability of the underlying model. The novelty is the workflow: a well-blocked take becomes a template. Shoot it once with one performer in one location, then recast it into as many worlds as you have reference images for.
Motion Transfer vs Object Swap
Two modes, and the difference between them decides which one you want. It comes down to how much of the shot you are willing to change.
| Motion Transfer | Object Swap | |
|---|---|---|
| What you change | The whole world: cast, location, wardrobe, style | One element you point at |
| What stays | Camera path, framing, timing, the performance | Everything else in the frame |
| What you supply | Source clip plus references for the new scene | Source clip, the replacement, and a short description |
| Reach for it when | You want a different world entirely | You like the clip and want one thing changed |
The official use cases map cleanly onto that split. Style swap — turn an anime-style clip into live action or the reverse, keeping the motion, timing and camera work. Wardrobe change — swap outfits on the same performance by uploading different references. Ad variation — change the character, product or setting for a different market while keeping the pacing and structure of the original. Content repurposing — turn a raw phone clip into something that reads as a commercial without losing the original action and framing. AI influencer content — keep one underlying motion and change who is on screen.
Read those five again and a business case appears: one shoot, many deliverables. That is the entire pitch, and for an agency or a creator producing ad variants it is a real one. It is also why the pricing is set the way it is.
What a render costs
Genjutsu is not sold on its own. It draws on your Higgsfield credit balance, like everything else on the platform, and the charge scales with clip length and output resolution. Higgsfield's own table below is for a 15-second clip, verified September 2026:
| Resolution | Credits | Approx. price |
|---|---|---|
| 480p | 40 | $2.00 |
| 720p | 104 | $5.20 |
| 1080p | 144 | $7.20 |
Two things about this table matter more than the numbers. First, the jump from 480p to 720p is 2.6×, not the 1.5× you might expect — the middle tier is where most of the cost lives, so try the edit at 480p before you commit to 720p. Second, Higgsfield does not publish a fixed per-render price: the exact charge shows in the generator immediately before you submit, and it depends on the settings you chose. That is worth knowing because it means an old blog post quoting a different figure is not necessarily wrong — it is just a different configuration.
Where you can use it
- Inside Higgsfield — open the video tool, pick Genjutsu in the model selector. This is the only first-party route and the only one where Higgsfield's own prices and terms apply.
- Through Higgsfield's API — Higgsfield introduced its own API in September 2026, and third-party product write-ups list Genjutsu as reachable through it. If you are building rather than clicking, this is the path — but confirm the model list on the API documentation rather than on a comparison page.
- Through resellers — several platforms package a "keep the motion, recast the scene" edit under their own branding and run it on their own stack, sometimes on a different underlying model entirely. These are not Genjutsu, whatever the marketing says. They may still be cheaper, and some refund renders that fail. Judge them on their own terms, not on the Genjutsu name.
One practical point that catches people: because Genjutsu is a video tool rather than a generator, your clip length sets your price before you write anything. Trimming the source to the four or five seconds you actually want to recast is the single biggest cost decision you make, and you make it at upload time, not at the settings screen.
The line you should not cross
This section is the reason this page exists in the form it does. Genjutsu makes a specific kind of edit easy, and that edit has a specific kind of victim.
In September 2026 the recasting idea escaped the hip-hop and meme audience it started in and moved into sports. A real sideline reporter was targeted with AI clips of things she never did or said. The sequence, according to contemporaneous coverage, was: a genuine on-field interview moment became an AI clip of her being struck, a college quarterback publicly called it funny, apologised after the backlash, and she accepted the apology. Then a second wave followed — a further fabricated clip of the same woman, drawing roughly 1.8 million views in a day.
Three things are true at once here, and they are worth separating.
- The tool is not the problem. Higgsfield built a motion-transfer workflow and sells it with legitimate use cases: wardrobe tests, ad variants, style conversion. Nothing about that is improper.
- The output can be defamatory. Putting a real, named person into footage of actions they did not take is not a joke format, it is a false statement of fact about a specific human being, published to millions. Publicity rights, defamation and platform enforcement all point the same way.
- You will not win the argument on technical grounds. "It was AI" is not a defence when the whole point of the clip is that viewers cannot tell.
What Higgsfield does not publish
Three gaps you should know about before you plan around this tool.
| Question | What is actually on the record |
|---|---|
| Which model is underneath | Not published. Genjutsu names a feature, not an architecture. Community threads guess at Seedance 2.5 and several competing tools advertise running the same class of edit on Seedance, but Higgsfield has not confirmed it. The practical consequence: you cannot compare it on model quality, only on output and price. |
| Realistic processing time | Not published, and disputed. Community threads from September 2026 include multiple users reporting renders sitting in a queue for two to four days with no ETA. We could not verify those independently, and happy users do not post. Treat it as a signal to test one cheap generation before you schedule a campaign around it. |
| The exact input bounds | Inconsistent across Higgsfield's own pages. The changelog and product page say one video up to 30 seconds and up to 30 reference photos; the blog has used a 4-second floor while the product page has used 3; third-party coverage reports up to 40 reference images. Confirm in the generator rather than in an article, including this one. |
Also absent, as with every tool in this category: any independent, like-for-like benchmark. There is no third-party test putting Genjutsu head to head with a general video model doing the same recast. Which means the useful comparison is the only one you can run — your own clip, both routes, same reference images.
Getting a usable recast
Genjutsu rewards a specific kind of source footage and punishes everything else. The order below is roughly the order of how much each step affects the result.
-
Shoot for the transfer, not for the edit.
The whole promise is that the motion survives. That requires motion it can read: one performer, clear silhouette, stable framing, deliberate movement. Fast cuts, heavy motion blur, two people crossing, or a camera that never settles all make the extraction harder and the result worse. The take you would have reshot anyway is not the take to recast.
-
Cut the source down before you upload.
Price scales with clip length and the tool does not care what is in it. Trim to the seconds you actually want. This is also the cheapest fix for quality — a shorter, cleaner clip beats a longer one in both cost and coherence.
-
Choose the mode by asking how much you want to keep.
If the answer is "everything except who is standing there", that is Object Swap. If the answer is "the movement, but a different world entirely", that is Motion Transfer. Picking the wider mode when you wanted a narrow edit is the most common way to waste a render.
-
References should be boring and consistent.
Same person, same lighting direction, same framing, several angles. Reference images are what the model rebuilds the scene from, so a set that disagrees with itself produces a character that changes between frames. This is the step where you can most cheaply raise the odds of a usable first render.
-
Try the preset before you write prose.
There are 30 or more presets and a full prompt is not required. Presets exist precisely because a preset is unambiguously readable by the system and a paragraph of adjectives is not. Reach for a short description only to correct something the preset got wrong.
-
Rehearse at the cheapest resolution, publish at the highest.
Because 480p to 720p is a 2.6× step, getting the mode and the references right at 480p and only then spending 720p or 1080p credits on the final take is worth more than any other optimisation on this page.
-
Label it, and keep the permission.
Publish with the platform's AI marker. Keep written consent for every face in the frame and every reference image you uploaded. If a client or a platform ever asks, that file is the difference between a conversation and a takedown.
Frequently asked questions
What is Higgsfield Genjutsu?
A video-to-video tool inside Higgsfield, released 1 September 2026 per Higgsfield's own changelog. It takes a clip you already have and rebuilds part of it. Motion Transfer lifts the motion, camera work and timing out of your source and builds a new scene around them from your references. Object Swap changes one element you point at and leaves everything else as filmed. No full prompt needed — 30 or more presets, with an optional short description.
How much does it cost?
It runs on Higgsfield credits, not a separate price. Higgsfield's table, verified September 2026: a 15-second clip is 40 credits at 480p (~$2.00), 104 credits at 720p (~$5.20) and 144 credits at 1080p (~$7.20). The exact charge appears in the generator before you submit, so there is no fixed per-render figure — it depends on your clip length and resolution.
Is there a free way to use it?
No. The launch post said "try it for free", but Higgsfield's pricing page in October 2026 shows the free plan with zero credits and no commercial-use rights — you can see the interface but cannot render. Every paid tier unlocks commercial use. See what each plan actually grants.
Motion Transfer or Object Swap?
Ask how much you want to keep. Want everything except the person? Object Swap. Want the movement but an entirely different world? Motion Transfer. Choosing the broader mode for a narrow edit is the most common way to burn a render for nothing.
How long can the input be, and how many references?
Higgsfield's changelog and product page say one video up to 30 seconds and up to 30 reference photos, output up to 1080p. Some third-party coverage says 3 to 30 seconds and up to 40 references, and Higgsfield's own blog used a 4-second floor while its product page used 3. Verify in the generator on the day; one clip and one reference set per generation.
Which model does it run on?
Not published. Genjutsu is a Higgsfield-branded workflow rather than a named model — community threads guess at Seedance 2.5, and competitors advertise running similar edits on Seedance, but Higgsfield has not confirmed anything. Judge it on output and price, not on an assumed architecture.
Can I use the output commercially?
Higgsfield says you own what you generate and may publish it across organic and paid channels, subject to its Terms of Use, from Starter upwards — the free plan carries no commercial rights. Its terms also prohibit using outputs to train other AI models, and note that your uploads may be used to train Higgsfield's own — relevant before you upload footage of anyone else.
Is this what makes Hotel Lobby AI clips?
It is the tool most often named for that style of edit, and it is built for exactly that job. But the Hotel Lobby format predates it and is usually made from two reference photos with a general video model rather than by recasting footage. See the Hotel Lobby AI format for that workflow.
Can I use a photo of a real person as a reference?
Not without consent, and never to put words or actions into a public figure's mouth. In September 2026 the same technique was used against a real sideline reporter; one fabricated clip drew about 1.8 million views in a day. That is defamation and publicity-rights exposure, and platforms remove accounts for it. Use your own face, a licensed performer, or an original character. We do not publish the other kind of tutorial.
Also on this site
- Hotel Lobby AI — the orange-booth duet format: what it is, the five rules that make a clip read as the trend, eight prompts and the permission rules.
- Higgsfield AI review — the plan ladder behind Genjutsu: what each tier grants, why credits vanish faster than the allowance suggests, and which surfaces bill even on an unlimited model.
- Seedance 2.5 — the model most often guessed at as the engine underneath. What ByteDance published, what each route charges, and the four specifications nobody has confirmed.
- Wan 3.0 — Alibaba's 30-second model, and the first that accepts a PPT, PDF or spreadsheet as input.
- Image to video AI — the wider category, with the four-step workflow and the prompt pattern that stops faces morphing.