How to Use Higgsfield Genjutsu: A Real Guide to Reshooting a Video (2026)

Higgsfield shipped Genjutsu on September 1, 2026. It takes a video you already have and renders it again with something inside it changed: a different character, a different product, a different location, while the movement and the cut stay where they were. That is a different job from text-to-video, and it is worth knowing exactly what the tool does before you spend credits finding out.

Every fact about Higgsfield here comes from their own product page, blog, changelog and help center.

A disclosure before we start: we make Riffkit, which has its own reshoot mode, Swap. If you want a second version of a video you already made, or one that runs past Genjutsu's 30 seconds, Swap reshoots it at full length (currently up to 45 seconds) and changes only the person, product or detail you name. If the winner is someone else's, its Adapt mode keeps the beats and rebuilds it as a new video, which is the safer route on TikTok Shop.

What Genjutsu actually does

Genjutsu is video-to-video. You give it one clip, optionally some reference images, and it rebuilds that clip with the parts you name replaced. Higgsfield's changelog describes it as changing "an object, location, product, or character in a video while leaving everything else as it was." The product page frames it as swapping characters, outfits, locations and objects, or keeping only the motion and rebuilding the rest.

The word doing the work there is rebuilding. This is not a filter pasted over existing pixels. The output is a fresh render that inherits the original's movement, camera and cut points, which is why the failure modes look nothing like the ones you know from face-swap apps.

Access path, from the changelog: Video → model selector → choose Genjutsu. The marketing entry point is higgsfield.ai/genjutsu.

Motion Transfer vs Object Swap

There are two modes, and picking the wrong one costs you a render.

Motion Transfer keeps the movement and rebuilds the rest from your references. Higgsfield's FAQ: it "uses the motion from an existing video as a reference, then applies that movement to a new character or rebuilt scene." Reach for it when you want the performance and none of the original's look, such as a proven dance or camera move carrying your own creator in your own set.

Object Swap targets one element and leaves the surrounding shot alone. Their FAQ: it "replaces the object or element you specify while preserving the rest of the original footage." Reach for it when a demo already works and only the thing in frame is wrong, such as a competitor's bottle where your SKU should be.

The practical question is how much of the original you want to survive. Object Swap protects the most and asks the least of your references. Motion Transfer gives you a new world and puts far more weight on the reference images you supply.

What you need before you start

• One source video, 4 to 30 seconds. The product page says Genjutsu supports reference videos from 4 to 30 seconds, and the blog FAQ repeats it. The changelog describes the input as one video of up to 30 seconds. There is no documented multi-clip input, so anything longer has to be cut up and joined back together yourself.
• Reference images, if the change needs them. Up to 30 per generation. Higgsfield suggests using that headroom for cycling through several outfit or style options against the same clip.
• A preset, or a sentence. The changelog says there are 30+ presets, and the FAQ states that a preset alone can drive the entire change with a short description optional for anything more specific. Prompting is not the main interface here.
• Credits, and a paid plan if you want no watermark. Higgsfield's help center is blunt about this: on a free account every generation carries a Higgsfield watermark, and on any paid plan generations come without one. There is no watermark toggle.

Genjutsu is priced in Higgsfield credits per render; check Higgsfield's pricing page for the current rates.

Your first reshoot, step by step

Higgsfield's own five steps, with the parts that matter in practice.

1. Choose a feature. Motion Transfer or Object Swap, decided before anything else, because it changes what your reference images have to carry.

2. Upload your source video. Trim to the window you actually want first. Do not upload 30 seconds when 8 of them are the ones you care about; a clip trimmed to the part that matters is easier to judge, and the Generate button shows its cost before you confirm.

3. Add reference images if needed. This is where a swap is won or lost. For a character, give the same person from more than one angle in consistent lighting. If you are reusing a presenter across many videos, the reference set is the asset, not the render. We wrote up why characters drift and how to lock them in keeping an AI character consistent across videos.

4. Choose a preset or describe the change. Presets are the intended path. Write a description when the change is specific, and name the thing you want changed rather than describing the whole scene.

5. Generate and review. The credit cost appears on the Generate button before you confirm, so check it against the resolution you selected rather than assuming.

What it really costs

Genjutsu is priced in Higgsfield credits per render, and the cost for your settings appears on the Generate button before you confirm. Three things decide your real cost per usable shot:

Retries cost credits, as on any generator. The help center states that credits are deducted for all generations, including re-rolls. Nobody's first reshoot of a new format is the keeper, so budget the attempt count, not the render count.

Failures are usually refunded, but not guaranteed. The help center says the refund for a failed generation is automatic and lands within a few minutes, then adds that specific models might not refund credits for failures.

Subscription credits do not carry over. Higgsfield's own wording: they "reset regularly and don't roll over: on monthly plans with each paid renewal, on annual plans every 30 days." On an annual plan you are not banking a year of credits, you are spending a monthly allowance, and an unused balance is never a reserve you save for launch week.

Plan prices and promotions move, so read higgsfield.ai/pricing yourself rather than trusting any number a blog quotes you, including this one.

What the documentation does not answer

Worth knowing before you build a workflow on it, because we looked for each of these on Higgsfield's own pages and did not find them:

• Which model powers Genjutsu. Higgsfield does not name an engine anywhere we could find. Creators comparing it to specific models are guessing.
• Whether the source video's aspect ratio is preserved. Nothing states this either way, which matters if your source is 16:9 and your channel is 9:16.
• How audio is handled. No page says whether the original audio track is kept, regenerated or dropped. The launch post's list of what stays from the source covers motion, timing, camera work, pose, pacing and framing, and stops there.
• Whether a social link works as a source. Every official description says upload your video. There is no documented link import.
• What "many versions" buys you. The product page offers to "Run many versions from one setup, then generate more of whatever works," without saying how many one setup covers or what the extra ones cost.

Where a reshoot drifts

We built the same capability into Riffkit, so this section is what reshooting breaks on in our own renders, not a measurement of Genjutsu. Expect to meet at least some of it in any tool of this shape.

A face described in words alone will not hold. If the change is "make her a 30-something founder" with no reference image, the person can shift between shots. A reference-backed character is a different reliability class from a sentence, which is why our own product nudges you toward picking a character rather than describing one.

Reference images bring their whole frame with them. Wardrobe, lighting mood and background texture bleed in from a reference shot you only meant to use for a face. Clean, neutral references beat pretty ones.

Only what you name changes. Supporting cast, props at the edge of frame and background signage stay as they were unless the instruction points at them. That is usually the feature, and occasionally the bug when a background element is the thing that dates the footage.

Source audio decides where the video can be posted more often than the picture does. A commercial track in the original clip is a rights problem the render does not solve.

Before you post

Two checks worth building into the habit.

Originality, and the platform rule that outranks it. Reshooting someone else's footage is a different legal position from making your own video that follows the same structure; we covered that line in can you legally remake a viral TikTok, and the short version is that the formula travels and the footage does not. The stricter rule is the platform's: TikTok Shop's unoriginal-content policy names "Recreating a video that directly mirrors the actions of another creator's content", and also "Using AI-generated content that results in duplicate or highly similar videos to other creators". Generating fresh frames does not exempt you from it. If you sell on TikTok Shop, reshoot your own videos, and rebuild someone else's winner as a new video rather than mirroring its shots.

The AI label. TikTok's disclosure toggle is off by default, sits somewhere you will not find by accident, and cannot be added after posting. Here is where it lives and when to use it.

If the video you want to reshoot is one you already made

Genjutsu starts from an upload every time. Swap, the Riffkit mode from the disclosure above, does not: it reshoots window by window on the source's own clock, keeping the shots, cuts, pacing, sound and aspect ratio. The source can be a template, a new upload, a TikTok link, or one of your own finished Adapt or Create renders, picked from your library with no re-upload. A video that came out of Swap is not accepted back as a source. Swap refuses outright if you have not named a change, and shows a per-second rate for every engine before anything renders, so you can price a 9 second job and a 40 second one first.

For how per-second pricing compares across engines, see Seedance 2.5 vs MiniMax H3 real costs. For a wider look at the tools in this category, read Higgsfield Genjutsu alternative.