HeyGen announces Avatar V

Background

An AI avatar model takes a recording of a person and generates new video of that person: saying lines they never recorded, appearing in settings they were never filmed in. The hard part is usually not a single frame. A face that looks convincing in a still can begin to slip once the video runs — the likeness softens, the movement stops matching the voice, and the viewer registers, without being able to point at a moment, that the person on screen has stopped being quite the same person. That slow loss of identity is the problem HeyGen says its new model was built to address ^[1].

What happened

On 8 April 2026, HeyGen announced Avatar V, describing it as the most realistic AI avatar model in the world ^[1].

Alongside the model, HeyGen described a change in how it is used. Avatar V is now the foundation that the rest of HeyGen runs on ^[1]. In other words, it is presented not as an extra tool offered next to earlier ones, but as the base layer underneath them.

Key details

The input side is deliberately short. One recording of about 15 seconds is enough to generate studio-quality video that maintains the user's face, voice and presence ^[1].

HeyGen frames the goal as identity consistency across angles, looks and runtime, rather than as a single impressive frame ^[1]. The phrasing matters. Consistency across angles and looks is about holding a likeness when the viewpoint or the styling changes; consistency across runtime is about holding it for as long as the video keeps playing.

The company also names the specific failure it is aiming at. It frames the problem as the uncanny drift that appears roughly twenty seconds into avatar videos ^[1] — the point at which the output stops reading as the same person. Defining the target that way turns realism into something measured in seconds of continuous video rather than in a single still.

The announcement also pairs Avatar V with a second model: it is designed to be paired with Seedance 2.0 for cinematic footage ^[1].

Why it matters

Three aspects of the announcement are worth separating out.

The first is where the quality bar is set. By talking about drift at around twenty seconds and about consistency across runtime, HeyGen describes the model's performance in terms of how long a generated person stays recognisably themselves, not how good one frame looks ^[1]. For anyone using avatars to deliver several minutes of spoken content, duration is the property that decides whether the output is usable.

The second is the cost of getting started. A recording of roughly 15 seconds is short enough to be made on a phone in a few minutes, and the company says it is sufficient to produce studio-quality video preserving face, voice and presence ^[1]. The shorter the required input, the fewer reasons there are to postpone making an avatar at all.

The third is the model's position inside HeyGen. Being the foundation that the rest of HeyGen runs on means improvements to Avatar V are not confined to one product; they propagate through whatever else the company offers ^[1]. Pairing it with Seedance 2.0 for cinematic footage extends the same logic to a second model, so that the avatar supplies the person and the companion model supplies the kind of footage the person appears in ^[1].

What the announcement leaves open is how the model behaves beyond the cases HeyGen describes. The targets it sets — the most realistic avatar model, studio-quality output from a 15-second recording, a likeness that holds past the twenty-second mark — are targets rather than measured results ^[1]. They are nonetheless specific ones, and they say clearly what Avatar V is meant to be: a base layer for HeyGen's other products, and a person who stays the same person for longer than before.