How to Turn an AI Character Into a Video (Step by Step)

From a Still Character to a Clip That Holds the Same Face

To turn an AI character into a video, save the character first, then call it by its tag inside a video model that supports Characters. The character carries its own identity into the clip, so you describe the motion and the model holds the face. It runs in a browser, and a first clip takes minutes rather than hours: about 2 to 3 minutes in Fast Mode, and up to around 20 on the Cherry models without it.

Holding the face is the whole trick. Faces wobble in motion in ways they never do in stills, and no amount of prompt detail fixes it, because a description of a person isn't the same thing as a saved person.

This walkthrough covers the loop end to end: what to set up, how to write the shot, why a face drifts anyway, and what to do about it.

What You Need Before You Start

2 things: a character you've already saved, and a video model that supports Characters. If you don't have the character yet, make it first. One clear photo is enough, and the full process is in how to create a consistent AI character from one photo.

The feature doing the work is Character Reference (Image)-to-Video: the model reads identity from your saved character rather than from the words in your prompt. That's the distinction to get right, because it's where most people lose the face.

The lookalike to watch for is first-and-last-frame image-to-video. It anchors the opening and closing frame of a shot, and it has no idea who your character is. Animate a character that way and you're asking the model to guess the face in every frame between the 2 you gave it, and it will.

How to Turn Your Character into a Video (Step by Step)

• Open your character and check the tag. Pull up the character you want to animate. The name you gave it is the tag you'll type, so confirm the spelling. If the tag doesn't match a saved character, the prompt is just describing a person instead of calling yours.

• Switch to a video model that supports Characters. Mage's fruit video models that support it are Cherry, Cherry Mini, Cherry Pro, Melon, Berry 2, Raspberry, Berry, and Plum. Any of them will hold your face, and the table below covers what each one is suited to.

• Write the shot around the tag. The tag carries the face, so everything else in the prompt is the shot: what the character does, how the camera moves, what the light is doing. Motion first, then framing, then light.

• Set duration and resolution before you generate. Shorter is steadier. 3 or 4 seconds gives identity fewer frames to drift across, and you can stack short clips into a longer sequence later.

• Generate, then cut the misses. Each generation varies. Run the shot a few times, keep the takes where the face holds all the way through, and bin the rest.

A working prompt looks like this:

@maya turning to look over her shoulder, slow push in, warm streetlight, shallow depth of field

Keep the motion plausible. A character walking, turning, laughing, or reaching for something holds far better than one doing a backflip, because the model has to invent less of the face on the way through.

If a take is close but the face softens near the end, shorten the clip and regenerate. That fixes more takes than rewriting the prompt does.

Why the Face Still Wobbles in Motion

That failure has a name: identity drift, the face sliding off the person you locked. Video is where it bites hardest, because a still only asks the model to match one pose, while a clip has to hold the same bone structure across every frame as the head turns, the mouth moves, and the light travels across the face.

Most wobbles trace back to 1 of 4 things:

• A soft reference. The character is only as sharp as the photo it was locked from. A blurry, filtered, or steeply angled source produces a face that was already vague before motion entered the picture.

• A small face in frame. When the face occupies less than a meaningful portion of the frame, there aren't enough pixels to keep the features specific. Generate the close shot, then build wider coverage around it.

• Too much motion. Fast turns, whipping camera moves, and heavy occlusion all force the model to rebuild the face from scratch mid-clip.

• A clip that runs long. Drift compounds. 4 seconds that hold beat 10 that fall apart at second 6.

A waxy, over-smoothed face is a different failure from a drifting one, though the two are related: fine skin and eye texture is part of what makes a face identifiable, so smoothing it away costs some of the likeness even when the bone structure holds. On a still, Skin Enhancer adds skin texture and detail back into a portrait. For a finished clip, Video Enhancer is the app that upscales and refines video. The deeper mechanics of why identity slips at all are in AI Character Consistency: Why Faces Drift and How to Lock Them.

Which Models Hold a Character in Video

Any of these will carry a saved character. The Characters system handles the face on all of them, so pick on clip length and output quality instead. For a wider comparison across the field, see our roundup of the Best AI Video Generators for Consistent Characters in 2026.

Gems lift the resolution ceilings: Berry 2 reaches 1080p, Cherry and Cherry Mini top out at 720p, and Cherry Pro and Plum go to 4K. Clip length is not the differentiator, since every model in the table reaches at least 15 seconds on Gems. So you can test a shot on the highest-quality option before deciding which tier you actually want.

One caution that matters more than the choice between them: models that run on first and last frames only, without Characters or References support, won't hold your character. Check for Characters support before you commit a shot to one.

Give the Character a Voice, Then Stack the Clips

Once the character animates cleanly, 4 features extend what you can do with it:

• Character Voice, still in beta, lets you upload a voice file and reuse it so a saved character sounds the same across video, instead of like 3 different people across a series. It works on Cherry, Cherry Pro, Cherry Mini, Raspberry, and Plum.

• Audio Generation, powered by Seed Audio, generates character voices, music, and sound effects from a text prompt, reference audio, or a reference image. It's unlimited for all Premium members.

• Multi-Characters, also in beta, puts multiple consistent characters in the same creation. Tag each one, and they stay distinct instead of blending into each other. Mage documents it on the Mango V2 image model, so for a two-character video scene the Scene Builder app is the built-for-it path: add 2 characters, pick a scene preset, and adjust it from there.

• Motion Control takes any video as a motion reference and transfers that movement onto your character, which is how you get a specific performance instead of whatever motion the model invents.

When the individual clips are good, Storyboard sequences up to 6 finished clips into one continuous story, so a set of 4-second takes becomes a scene.

Tips for Clips That Hold

• Lock the character before you shoot anything. Get the face right in stills first. Animating a character you're not happy with just adds motion to the problem.

• Change one thing per take. New motion, or new lighting, or new framing. Not all 3, or you won't know what broke the face.

• Start close, then go wide. Close and mid shots hold identity best, so build those first and use them as your reference for the wider coverage.

• Keep the tag identical every time. The tag is what carries the face, so a renamed or misspelled one pulls in nothing.

• Cut long shots into short takes. 3 takes of 4 seconds are easier to keep consistent than 1 take of 12, and Storyboard puts them back together.

Frequently Asked Questions

How do I turn a photo into a video with the same face?

Save the photo as a character first, so the face becomes a reusable identity rather than a one-off image. Then call that character by its tag in a video model that supports Characters and describe the motion you want. Going straight from photo to video without saving a character is what produces a face that drifts.

How do I keep a character's face the same across a whole AI video?

4 things do most of the work: use Character Reference (Image)-to-Video so the saved character supplies the identity, keep the face reasonably large in frame, hold the clip to a few seconds, and keep the motion plausible.

Can I animate an AI character without training a model?

Yes. Training a Low-Rank Adaptation (LoRA) is the traditional route and still earns its keep on very high-volume work, where a reusable custom model pays back the effort. But locking a character from one image and carrying it into a clip needs no dataset and no training time.

Why does my character's face change halfway through the clip?

Usually the clip is too long, the face is too small in frame, or the motion is too aggressive. Shorten the take, move the camera closer, and calm the movement. If it still slips, the original reference photo is the thing to improve.

Do I need a powerful computer to make AI character videos?

No. It runs in the browser, with no graphics processing unit (GPU) of your own and no setup.

Can I put 2 characters in the same video?

Yes. The Scene Builder app is purpose-built for it: add 2 characters, choose a scene preset, and generate, on Cherry, Cherry Pro, or Raspberry. In a normal prompt, tag each character by name so each keeps its own face, since a model asked for 2 people without tags tends to blend them.

Does the character have to be a real person?

No, and it often shouldn't be. Mage prohibits non-consensual imagery of real people, so the line is your own face, a face you have explicit permission to use, or an original one. An original face is the cleanest build.

Shoot Your First Character Clip

Pick a character you've already locked, write one close shot with a single simple movement, and run it at 4 seconds. That one clip tells you more about how far your reference will carry than another afternoon of still images will.

Then go wider, add a voice, and stack the takes into a scene. Have fun creating!