HeyGen Avatar IV: make a portrait talk with your own audio

One still portrait, one voice recording, and a short render later the face is talking. That is the whole promise of HeyGen Avatar IV, and it is a good one: a character for a game trailer, a narrator for a course, a painted figure that introduces your exhibition. It also has a text field for directions, which most talking-head models lack. Most disappointing results come from the same handful of mistakes, so this guide is built around them.

Mistake 1: a portrait that hides the mouth

The model has to animate lips, jaw and eyes from one picture. Give it a face that looks roughly at the camera, with the mouth fully visible, even light and nothing in front of it: no hand on the chin, no microphone, no hair across the lips. A head and shoulders crop works better than a full-body shot where the face is a few pixels wide. Generated portraits work too; if yours looks waxy, fix the skin first with the tips in the plastic skin guide, because the animation will not.

Mistake 2: leaving the directions empty

The prompt is optional, and that is exactly why people skip it. Without it, the face does what the audio suggests. With it, you can ask for a mood, a habit, a gesture. Write it like notes to an actor: how the person feels, where they look, what the head does, and what they should not do.

Mistake 3: turning expressiveness all the way up

Expressiveness sets how much the face and head move: low (the default), medium or high. High is fun for a cartoon or a game hero and too much for a calm explainer, where it can look like the person cannot sit still. Real faces in photos usually look best on low or medium; save high for characters whose style already exaggerates.

Mistake 4: forcing an aspect ratio that fights the photo

Aspect ratio is optional here. Left empty, the model uses the supported ratio closest to your image, which keeps the framing you chose. Set it (16:9, 9:16, 1:1, 5:4 or 4:5) only when the destination needs it, and then crop the portrait to that shape yourself before uploading, so you decide what gets cut, not the model. A background image is available under the advanced settings if the original backdrop does not fit.

Mistake 5: using a face you have no right to use

A talking video of a real person saying something they never said can hurt them, and the terms forbid deepfakes of a real, identifiable person. Use your own face, a fictional character you generated, or a portrait you have clear permission for. For animating a photo without speech, turning a photo into a video is the gentler route.

When to pick something else

Avatar IV is about the face and the voice. It will not walk the person through a scene or change their clothes, and it tops out at 1080p. If you do not have a portrait at all and need a presenter in a real room, Avatar V uses ready-made avatars at up to 4K; the Avatar V guide explains it. Every other talking avatar and lipsync model is on the models page.