AI talking avatar and lip-sync guide for portraits

A talking avatar is a portrait that appears to speak. The useful job is narrow: a library welcome, a product aside, or a lesson intro of one or two sentences. This guide uses a fictional library assistant, Mara Chen, as a planning example. Her line is: Your hold is ready at the front desk.

On OmniArt you can design the portrait in the image workspace and generate the voice with a speech model in the audio workspace. The public model catalog has image models, video models, and speech models. It does not list a dedicated lip-sync or talking-avatar control that maps a script onto mouth shapes. Finish that alignment in an external lip-sync editor, then review the file at the crop you will publish. Background on the category sits in avatar notes, lip-sync notes, and photo avatar notes.

What a speaking portrait is for

Use a speaking portrait when the face carries the message and the shot stays close. Mara's line works because the viewer needs one fact and a place to look: the hold desk, not a tour of the building. Plan one thought, one face, and one framing. A shelf shot belongs in a separate clip.

Get consent before you animate a face

Animate a face you created, or a real person who agreed to this exact use. If Mara is based on a colleague, get a written yes for the portrait, the voice, and the places you will post the clip.

Write the script and choose the voice

Write for a single breath. Your hold is ready at the front desk lands cleanly. Read the line aloud. If you stumble, cut words before you generate.

Generate the voice in OmniArt's audio workspace and keep one preset for the whole piece. The MiniMax Speech 2.8 guide covers HD and Turbo takes and punctuation. Keep one preset from the test line through the final line, including the short pause at each end.

Run a sync test on one short line

Before any external tool, make a short image-to-video test of Mara with a small nod or a blink. That test checks light, shoulders, and whether the face stays Mara. It does not claim the mouth matches words. A video model that accepts an image can move the still. That motion is not a control that binds frames to your voice file.

Take the approved voice into a lip-sync editor outside OmniArt and sync only this one sentence. Play it at normal speed with the sound on. The mouth should open with the first word and settle when the phrase ends, with the jaw still attached to the face.

If the test fails, change one input. Use a clearer front-facing still, a slower read, or a cleaner audio file. For example, reject a three-quarter selfie with the chin tucked if it hides the lower lip: a sync tool may have to invent the missing mouth edge. This is a rejection criterion, not a result from a test we performed.

Hold the face still across takes

Mara needs the same eyes, hair part, and open collar in the still, the motion test, and the synced export. Reuse the approved image. The character guide applies here even though the performance is one sentence.

The approved still is eye level, with the library window behind her left shoulder and nothing across the mouth. Sunglasses, a hand at the chin, heavy blur, and a steep profile hide the shapes a lip-sync editor has to read.

Review the clip at its final framing

Review the exported file in the crop you will publish. For a 9:16 library post, put captions below the chin so they do not cross the mouth. Check that the eyes still meet the lens after the crop, that the voice has no room hiss, and that the on-screen words match the spoken line.

Watch once for the meaning and once for the mouth. A frame that looks sharp while paused can still feel late when the sentence starts. Keep any music under the voice. If the crop slices the collar, return to the approved still.

Getting started on OmniArt

Create Mara as an original front-facing portrait in the image workspace, with both eyes and the mouth unobstructed. Generate the one-sentence voice in the audio workspace and save that take. Use a video model only for a short motion test of the still. Finish lip sync in an external editor, then watch the final framing with captions in place before you publish.