HeyGen releases Avatar V, trained from a 15-second clip

HeyGen announced Avatar V on 17 September 2026, describing it as the model "everything else in HeyGen runs on now". According to the company, the model builds an avatar from a 15-second clip with "no professional camera setup, no studio lighting, no crew required", and supports 177+ languages and dialects.

HeyGen's stated emphasis is consistency rather than peak fidelity: the announcement claims multi-angle consistency without drift and long-form stability across a full video rather than a short clip. The company also says Avatar V is the first of its models to separate performance from appearance, so a user can change outfit, setting and avatar version while preserving the original motion. HeyGen cites a G2 rating as evidence of realism.

The announcement contains no benchmark numbers, latency figures or quantitative comparison with competing systems.

Why it matters for digital humans: Long-form drift — an avatar that holds up for fifteen seconds and falls apart over five minutes — is the failure mode that keeps this class of tool out of training libraries and support deployments, so a vendor foregrounding stability over headline realism is a meaningful shift in pitch. The separation of performance from appearance is the more interesting claim for anyone maintaining a content library, since it implies re-rendering a recorded take in a new wardrobe without a reshoot. All of it is company-reported and unaudited: there are no published metrics here, and the G2 rating is a user-review ranking, not a measurement.