The best AI model to swap a person in a video: what each one does, and a real side-by-side test

Kling, Seedance, and most of the models in the AI video model lists start from a still and invent the motion. To swap a person, you need the opposite: keep the motion, the camera, and the timing of a video you already have, and change only who is in it. A small group of models does that, and each one does it a different way.

On November 14, 2026, we took a 5 second talking clip of Camila, an AI person from Hook Studio's premade cast, and swapped her for Layla, another person from the same cast, on Genjutsu and on Genjutsu Light. Both clips are below, next to the original.

The swap models do one of three jobs. Pick the job first, then the model.

1. Swap the person, keep the video

You give the model a video and a photo of the new person. It rebuilds the person and keeps the motion, the camera, the setting, and the length. In Hook Studio, this is a recast.

• Genjutsu Object Swap (Higgsfield, August 31, 2026). Replaces the person, product, or object you name and keeps the rest of the shot. It takes a 4 to 30 second video and up to 30 reference images, and returns up to 1080p (Higgsfield).
• Genjutsu Light (Higgsfield calls it Genjutsu Lite, October 20, 2026). The cheaper version: 4 to 30 seconds, up to 10 reference images, 720p or 1080p. It applies the change on its own, with no mode to pick (Higgsfield).
• Wan Animate, replacement mode (Alibaba's Tongyi Lab, September 2025). Puts your character into the reference video and matches its lighting. Its weights are open, so you can run it on your own hardware (arXiv, TechNode).

2. Move your character with someone else's motion

Here the video lends only its motion. The character comes from your photo.

• Kling Motion Control (Kuaishou). Kling 3.0 Motion Control reached Kling's API on March 4, 2026. It moves the character in your image with the motion of a reference video. The reference video can run up to 30 seconds when the character takes the video's orientation, and up to 10 seconds when it keeps the image's (Kling API docs).
• Genjutsu Motion Transfer (Higgsfield). The motion of a video you have drives a new character or a rebuilt scene, with 20 ready styles or a style image of your own (Higgsfield).
• Wan Animate, animation mode (Alibaba). Your character follows the moves of a driving video. A second version, Wan-Animate-2, was reported in August 2026 under the Apache 2.0 license, with several characters at once (ComfyUI Wiki, OpenSourceForU).

3. Edit the video from one example frame

• Runway Aleph 2.0 (Runway, May 21, 2026). You show it one frame with the change, preview the result as an image, and it applies the change across the clip, up to 30 seconds at 1080p. It is on all of Runway's paid plans, on the desktop web. Through Runway's API it costs $0.28 a second, with a minimum of $0.56 a generation (Runway).

Why not just use Kling or Seedance?

In Hook Studio, a Kling or Seedance clip starts from a still, and the model invents the motion. That is the right tool for a new shot. When the moves, the timing, and the camera of a video you already have are the point, a swap model keeps them. Higgsfield does not say which model Genjutsu is built on, and no source we found names one.

The source is a 5 second clip we made in Hook Studio on Kling 3. Camila sits at a café table and says: "Same photo, same line, different AI model. Which one looks real?" We recast it with Layla on Genjutsu Light and on Genjutsu, both at final quality, 720p, with the same reference photo. Nothing else changed. In Hook Studio, the agent quotes both models before a recast starts, and you pick one. The source passed Hook Studio's realism check. Recasts skip that check, so the labels on the two recasts below are our own read.

This is one take on each model, a sample and not a verdict. Hook Studio offers Genjutsu as the closest match and Light at about a third of the price. On this clip, the cheap one won. So try Light first, and pay for Genjutsu when the Light take misses.

The voice stays, and changing it is a separate step

Both takes kept the sound of the source, so Layla talks in Camila's voice. A swap changes who you see, not who you hear. To give the new person their own voice, mute the clip and record a voiceover, or make a voice line and lip-sync the recast to it. The steps and prices are in how to recreate a viral video with AI.

Check the take yourself

In Hook Studio, every still and clip goes through an automatic realism check. Recasts do not, so the check is yours. Look at four things:

• The face against the reference. Put a frame next to the photo you gave the model. Our Genjutsu take drifted here.
• The hands against the source. Play both clips side by side and watch the hands. A gesture the source does not have is motion the model made up.
• The end. Our Genjutsu take came out 0.3 s shorter than the source. Check that the last word and the last beat are still there.
• The clothes. A recast replaces the whole person, clothes included. Both takes dressed Layla in the white shirt from her photo, so pick a reference in the outfit you want on screen.

• Swap only footage you own or have the right to use: your own videos, UGC you bought when the agreement allows it, or a video whose creator said yes in writing.
• Label it. Turn on the platform's AI label when you post a video with an AI-made person in it.
• Do not post someone else's voice as yours. A swap keeps the sound of the source, voice included.

• A video of yours worked, and you want it with another person: a swap. Try Genjutsu Light first, then Genjutsu if the Light take misses.
• Your character should do the moves from another video, in its own look: a motion transfer model, such as Kling Motion Control, Genjutsu Motion Transfer, or Wan Animate.
• You run your own GPU and want open weights: Wan Animate.
• You want to change something in footage you shot, from one example frame: Runway Aleph 2.0.
• The hook is an idea, not a move: skip the swap. A still and a 5 second Kling 3 clip cost 107 credits in Hook Studio.
• Keep a swap to one take, up to 15 seconds in Hook Studio. Two recasts joined together are two takes, and the person's hair and look change between them.

For the models that make new clips from a still, see which AI video model to use for each job, realism vs control, and who makes Kling, Seedance, and Veo.