How the Genjutsu AI examples were made
The footage, reference images, models, prompts and credit costs behind every example in our library — and what we learned making them.
Every clip in the example library was rendered in this editor, with the same models, settings and prices you get. We didn't retouch the results; the only processing afterwards was re-encoding them for the web and labelling them “AI-generated”. Each example's prompt, reference images and credit cost are listed on its card, and Recreate loads them into the editor so you can check the result yourself.
The source footage
A Genjutsu AI render is only as good as the clip it starts from, and we can only publish clips we have the rights to. The library uses three kinds of footage:
- Free-licence stock. Clips from Pexels and the Mixkit Free Licence, both of which allow commercial use and modification. Mixkit also has a Restricted Licence that is personal-use only. Four of our first-choice Mixkit clips turned out to be under it, so we dropped them.
- AI-generated clips. When we couldn't find a clean, licensed shot — a runway walk, a corridor, a car on a sunset road, a man on a street — we generated one with a text-to-video model (MiniMax H3). These are fictional people and places.
- A royalty-free test clip from samplelib.com, used for the park pan.
Every source is a 4–5 second cut, the length most people start with. Standard sources are 720p, because Standard Swap and Motion transfer need at least 720p; the Pro examples render at 480p.
The reference images
References are the photos of what goes into the shot. All of ours are AI-generated, fictional characters and products: the arctic fox, the corgi in a silver gown, the cat in a leather jacket, the mint-green bubble car, the gold watch, the chrome puffer jacket and the robot.
The biggest lesson here was contrast. Our first reference drafts were close to the originals — a slightly different car, a similar watch — and the before-and-after would have looked flat. The final ones differ in colour and shape at a glance: a white sports car becomes a mint-green 1950s bubble car, a red-strapped smartwatch becomes a gold diamond watch, a denim shirt becomes a mirror-chrome puffer. The edit is just as hard for the model either way, but a viewer sees it immediately.
Models and settings
| Example | Effect | Mode | Length | Credits |
|---|---|---|---|---|
| Neon dancer → 3D arctic fox | Swap | Pro, 480p | 5 s | 250 |
| Same dance → rainy Tokyo street | Motion transfer | Pro, 480p | 5 s | 250 |
| Street dance → robot dancer | Motion transfer | Pro, 480p | 5 s | 250 |
| Boardwalk skaters → knitted yarn world | Restyle | Pro, 480p | 5 s | 300 |
| Park pan → hand-painted anime | Restyle | Pro, 480p | 4 s | 240 |
| Sports car → 1950s bubble car | Swap | Standard (Kling O3) | 5 s | 125 |
| Smartwatch → gold diamond watch | Swap | Standard (Kling O3) | 5 s | 125 |
| Denim shirt → chrome puffer jacket | Swap | Standard (Kling O3) | 5 s | 125 |
| Runway walk → corgi in couture | Motion transfer | Standard (Kling O3) | 5 s | 125 |
| Corridor walk → cat in a leather jacket | Motion transfer | Standard (Kling O3) | 5 s | 125 |
The whole library cost 1,915 credits to render, a little more than the 1,790 in one Pro pack. Pro runs Higgsfield's Genjutsu model; Standard runs Kling O3 for Swap and Motion transfer and FLUX.3 Edit for Restyle at a half to a sixth of the price. The Standard examples show what the cheaper mode can do when the source clip is clean.
The prompts
Almost every prompt follows the same two-part pattern: name the one change, then list what must stay.
Replace the white car with the mint-green bubble car from the reference image, driving the same line around the bend with its headlights on. Keep the road, the sunset backlight, the trees and the camera move.
The second sentence matters as much as the first. Anything a prompt leaves open, the model is free to rewrite — so the prompts spell out the face, the background, the lighting and the camera whenever they should survive. Restyle prompts are shorter because the preset carries the look; they only say what to protect (“Keep the slow camera pan.”).
Re-renders
The four Standard examples on AI-generated footage were rendered twice. Their first versions used Mixkit clips that we later found were under the personal-use licence, so we generated new source clips and rendered again with the same prompts and references.
What we'd tell a first-time user
- Start from a steady clip with one clear subject. Busy crowds and fast whip-pans are where every video-to-video model struggles.
- Use 720p or larger for Standard. Swap and Motion transfer won't take smaller clips, and sharper sources give sharper results. If your clip is too long, too large or a .mov, the editor can trim and convert it in your browser before uploading.
- Pick references that look nothing like the original. You'll see the edit better, and so will your audience.
- Write what stays, not just what changes.
- Choose the model for your task. A 5-second Standard Swap costs 125 credits. Pro uses Higgsfield Genjutsu; choose it when you want that specific model. Compare results on your footage rather than assuming the more expensive mode will look better.
Want to try one? Open the example library, press Recreate on any card, and swap in your own clip or character. The $4.99 trial pack covers one 5-second Pro render or two Standard Swaps.