Higgsfield's Genjutsu Keeps the Acting and Camera, Swaps the Background

The easiest way to picture it is a puppet show. The puppeteer's hand movements stay the same, but the puppet on stage gets swapped out. The pace at which an actor walked, the moment they turned their head, the rhythm of the camera shake — all of that carries over unchanged from the original. Only what's visible on screen is different.

This has long been the most frustrating part of AI video editing. Fixing even one element you didn't like meant tweaking the prompt and regenerating the whole clip, which meant losing the acting and camera work you'd already gotten right.

Two Modes

According to Higgsfield, Genjutsu runs on two modes.

Motion Transfer extracts only the motion and camera work from the source video and applies them to a new subject or scene. Because timing and camera movement are preserved, it's particularly useful for videos built around choreography or a musical beat.

Object Swap, by contrast, targets just one part of the frame. It changes only the specified element — clothing, a product, a prop — and leaves the rest of the shot untouched.

Higgsfield says a prompt isn't strictly necessary. More than 30 presets are available, and picking one alone is enough to run the conversion.

In a post on X, the company summed it up this way: "Keep the performance, the camera, and the edit exactly as they were, and pick whatever world you want." That's essentially how it draws the line between what stays and what changes.

Inputs and Outputs

Category	Official Spec
Reference video	One clip, under 30 seconds
Reference images	Up to 30 images (people, products, clothing, etc.)
Prompt	Optional, presets alone are sufficient
Output	Up to 1080p
Source type	Both live-action footage and AI-generated video

Pricing runs on a credit system, and the company says the interface shows how many credits a given generation will cost before you run it.

Rigging a Phone to the Blender Viewport

A demo released the same week takes a slightly different angle. Using a Blender add-on, Higgsfield showed a screen linked in real time to the 3D viewport, and said the shake of a handheld phone was carried directly into the rendered result, one to one.

The add-on itself is a toolbar that floats over the Blender viewport, letting users generate scenes, meshes, rigs, textures, and video directly into the scene. Higgsfield lists seven functions: scene composition, 3D model generation, character animation, image generation, video generation, camera control, and asset management that collects past generations.

Installation is a matter of dragging a zip file into the Blender window, and the company says it supports Blender versions 4.2 through 5.1. Generation runs on Higgsfield's own servers, so a local GPU isn't required — but that also means if you lose your internet connection, you can only view results you've already generated.

Generated outputs pile up in an asset library next to the viewport, and users drag them from there into the scene. The company says generation costs are displayed on screen before you run a job.

Put the two releases side by side and the company's target position becomes clear: let a human craft the camera movement by hand, and have the AI render follow that movement exactly.

Who This Is Built For

The official page lists six use cases: music videos that need to stay locked to choreography and rhythm, producing multiple variations of an ad that's already performing well, fixing a cut where only one element is off, maintaining a consistent character for an AI influencer, and regional localization that swaps out people and places.

The sixth is about rights. Higgsfield says outputs can be used in both organic content and paid advertising, as long as it's within the terms of service. According to the company, no video editing experience is needed — you just upload the video and reference images, pick a mode, and describe what you want changed.

Editor's Take

Any team that has actually tried putting AI video into production will spot the value of this tool right away. It's not about generation quality — it's about the cost of rework.

When you're producing an ad, the real time sink isn't generating the first cut — it's everything that comes after. Requests pile up: swap the model, the product packaging is outdated, this country's version needs a different background. Until now, every one of those requests meant starting over from scratch, and every regeneration wiped out the acting and camera work that had already been approved. Object Swap is aimed squarely at that pain point.

There's a specific friction point in the Korean market, too. Domestic ad campaigns tend to run on tight model contract periods and usage windows, which makes it hard to keep reusing the same footage for long. A feature that keeps the motion but swaps the person runs straight into that constraint. What the technology can do and what a contract permits are two different things, so only teams that have already sorted out portrait rights and contract terms will be able to actually use this tool.

What you need to prepare isn't equipment — it's material. Since the tool accepts up to 30 reference images, teams that already have model and product shots organized by angle have a built-in advantage. Cuts that used to get thrown away during a shoot become useful material here.

Last month, Higgsfield rolled out Layers, which automatically decomposes poster layers, and even released a 110-minute AI feature film. Taken together with Genjutsu, the company is clearly shifting its weight away from tools that generate new video and toward tools that rework video that's already been shot.