Keep a Consistent AI Character Across Images and Video

By the end of this post you'll have a repeatable workflow for a consistent AI character: one approved reference set, reused across new images, then carried into video. You'll also know where it still breaks, so you don't burn credits on the wrong fix.

Why characters drift

A text prompt describes a person. It doesn't identify one. "A woman in her 30s with short black hair and a denim jacket" matches millions of faces, so every run gives you a different one. Re-typing a longer description only narrows the pool a little.

The fix is to stop describing the face and start showing it. Current image and video models accept reference images, and that's what actually holds identity steady.

Step 1: Make one strong reference

Everything downstream copies your reference, including its flaws. Spend most of your effort here.

• Generate a clean, front-facing, evenly lit image. No hard shadows across the face, no sunglasses, no hand covering the mouth.
• Keep the background plain. Busy scenes leak into later results.
• Pick one outfit and one hairstyle you're happy to repeat.
• Add one or two extra angles (three-quarter view, profile) if you can. Runware's guide for Nano Banana 2 says to add separate references for the different viewing angles a scene will reveal and to use clean, well-lit references.

Reject any image where the face is slightly off. A "close enough" reference becomes a "wrong" character by image ten.

Step 2: Reuse the reference, don't re-describe it

This is the habit that changes results most. Runware's guidance puts it plainly: spend the prompt on the new scene, pose, lighting or medium, not on re-describing a face you've already supplied. That guide also says the model accepts between 1 and 14 reference images per request.

So instead of this:

A woman with short black hair, brown eyes, small nose, freckles, denim jacket, sitting in a cafe

Try this, with your reference attached:

The same woman from the reference image, sitting at a cafe window with a flat white coffee, morning light, candid photo, 35mm look

A few things help:

• Name what must stay. If the jacket or a tattoo matters, say so: "keep the denim jacket and the small star tattoo on her wrist."
• Change one thing at a time. New scene, or new outfit, or new style. Changing all three invites drift.
• Add a second reference for products or props. The same guide describes stacking a character reference with a product reference in one image.

The same approach can carry identity across mediums. Per Runware, one character reference can produce photography, watercolor or 3D animation without losing the likeness. In practice, expect style changes to loosen the match more than scene changes do.

Step 3: Use editing-style models for small changes

Some models are built to edit an existing image rather than generate from scratch. Black Forest Labs' FLUX.1 Kontext family was introduced for this, aimed at keeping characters consistent across edited images. The Batch's write-up is honest about the catch: these models "remain finicky, sometimes changing minute details or entire characters and backgrounds."

Take that seriously. Edits are the right tool for "same photo, different expression." They're a worse tool for "completely new scene," where a fresh generation from your reference often works better. And always compare against your original, not against the previous edit, so small errors don't compound.

Step 4: Carry the character into video

Video is harder, because identity has to hold across many frames and camera angles. There are two routes.

Route A: Start from your still

Use image-to-video with your approved image as the first frame. The face begins correct. Keep clips short and the motion simple, since a long turn of the head or a big camera move gives the model room to invent.

Route B: Reference images for the video model

Some video models take references directly. Google says Veo 3.1 lets you provide up to 3 reference images of a character, object or scene, which it describes as helpful for keeping a character consistent across multiple shots. Google's post doesn't promise a perfect match, and I wouldn't either. Treat it as a strong nudge, not a lock.

A prompt that works with either route:

The woman from the reference walks toward the camera along a quiet street, steady handheld shot, she smiles briefly. Same denim jacket.

Notice it describes the action and leaves the face alone.

A quick checklist before you generate

1. Is the reference sharp, well lit and front-facing?
2. Does the prompt describe the change, not the face?
3. Am I altering only one major thing this time?
4. For video: is the first frame my approved still, and is the motion modest?
5. Have I checked the result against the original reference, not the last output?

Common mistakes

• Using a stylized, moody reference. Dramatic lighting gets baked in. Shoot the reference neutral, then add mood in the prompt.
• Stacking too many references. More isn't always better. Conflicting angles or outfits confuse the result. Start with one or two.
• Judging from thumbnails. Eye shape, brow and jawline drift at full size. Zoom in.
• Expecting identical output each run. Results vary run to run. Generate a few options and keep the best, rather than rewriting the prompt after one miss.
• Changing models mid-project. Each model reads references differently. Pick one image model and one video model for a series if you can.

Where this still falls short

Be realistic. Small details like freckles, jewelry, tattoos and logos on clothing are the first things to slip. Extreme angles (from below, from behind) are weaker than front and three-quarter views. Crowded scenes with several consistent characters are harder than a solo subject. Google's Nano Banana 2 materials describe support for multiple characters, but you should test it with your own faces before planning a project around it.

Doing this in Mediafy Studio

If you'd rather not manage reference files by hand, Mediafy Studio has a couple of tools built for this. You can design an AI character once and reuse the same face (and voice) in images and videos, and type @ in the Studio to bring in a saved character. Seeds store a set of reference images for a consistent look. To be clear, Seeds are saved references, not trained custom models.

For a full chain, Flows lets you wire image, then video, then voice-over on one canvas and run it together. And if you make talking-head ads, UGC Factory takes a character and a product and produces a selfie-style image, lip-synced video and subtitles.

Credit costs show on the Generate button before you run anything. Run a quick draft on a cheaper model such as Seedream 5 Pro (2 credits) to test your composition before spending on a pricier model. If you want a guided start, the first image tutorial covers the basics, and the AI Creator Lab course goes deeper.

Ready to build your own character? Sign in with Google and try Mediafy Studio.

Sources

• Google, "Introducing Veo 3.1 and new creative capabilities in the Gemini API", Google Developers Blog: https://developers.googleblog.com/en/introducing-veo-3-1-and-new-creative-controls-in-the-gemini-api
• Runware, "Nano Banana 2: Character Consistency" guide, Runware docs: https://runware.ai/docs/models/google-nano-banana-2/guides/character-consistency
• DeepLearning.AI, "Black Forest Labs launches FLUX.1 Kontext for generating and altering images with consistent details", The Batch: https://www.deeplearning.ai/the-batch/black-forest-labs-launches-flux-1-kontext-for-generating-and-alterating-images-with-consistent-details