I Tested Genjutsu, a Video-to-Video AI: I Swapped the Actor and Checked If the Edit Survived

What do you do when the shoot is finished and you need to change the actor? Until now there was one answer: reshoot. New set, new lighting, new schedule, new budget.

This week I tested Higgsfield’s new video-to-video AI model, Genjutsu — not with their sample clips, but with my own footage. I swapped the person, the clothing and the environment, then checked whether the three things that actually matter survived.

What makes video-to-video different from normal AI video generation

Most AI video generation tools build something new from a prompt or a still image. Genjutsu does a different job: it takes footage you already shot and changes what is inside it, while preserving the decisions that made the shot work.

That distinction matters more than it sounds. A video is not a stack of frames. Camera motion, handheld shake, masks, layer order, the rhythm of the edit — they carry meaning together. Break one of them and the output immediately reads as machine-made.

The three tests that decide whether a model is usable

1. Does camera motion survive?

Fast pans and the natural instability of handheld footage came through unchanged. That is the Motion Control side doing its job.

2. Do masks and layer order hold?

The hardest test in my clip: a person walking behind an object. This is exactly where most models put the new character in front. Genjutsu kept the occlusion correct.

3. Character consistency across cuts

A classic failure is the character mutating into a different person at every cut point. I tested this specifically at 06:45 in the video — the character held across the edit.

Why this matters if you sell online

The most practical use case is at 09:45: producing localised ads for different markets from a single shoot. Film the product once, then adapt the presenter and the environment per market. Production cost has always been the wall in front of small brands, and this is precisely what AI content creation is starting to remove.

Second practical case, at 08:15: changing only the clothing. You update one detail after the shoot without reopening the whole production.

The limits, honestly

The model is carrying a lot at once, and in some frames the background behaves differently than expected. I run that reality check at 10:30 in the video. So no, not every output is flawless — but this is the first model I have used that does not destroy the edit in the process.

Try it

Higgsfield is currently offering Genjutsu with unlimited generations for 7 days at 70% off:

Explore Genjutsu →

Disclosure: This content was produced as part of a paid collaboration with Higgsfield (#ad). The link above is a tracking link; if you buy through it I may earn a commission at no extra cost to you. This article is educational and based on personal testing; it does not guarantee results. All footage in the video is my own.