Breaking

HeyGen Says It Generates 30-Minute AI Avatar Video in a Single Pass

June 30, 2026 at 12:43 EDT

Media Generation
Foundation Models
AI Agents

HeyGen has announced that its avatar generation system, "HeyGen Models," can produce a 30-minute AI video in a single continuous inference pass. The company frames this as six times the industry ceiling, placing its technical emphasis not on length but on consistency.

HeyGen Models · Avatar Inference at Scale

One Pass, 30 Minutes: AI Avatar Video Without the Drift

HeyGen says its streaming pipeline can render a 30-minute avatar video in a single continuous inference pass — six times the industry ceiling — with the real story being consistency, not length.

30 min

Single continuous inference pass

<5 s

Time-to-first-frame (Avatar IV)

27+ FPS

Frame rate at 720p

6× the industry ceiling

Single-pass length vs. the typical limit of short-clip generators.

1×

Typical
ceiling (~5 min)

6×

HeyGen
30 min

How the streaming pipeline holds together

Audio → DiT

Diffusion Transformer + flow matching drives video from audio

→

Chunk stream

N-chunks + interpolating I-chunks keep boundaries continuous

→

SR → VAE out

Identity-aware super-resolution, streamed VAE frame decoding

Constant GPU memory — usage scales with chunk size, not video duration. No fixed length limit; open-ended generation becomes possible.

The differentiator: no drift

Face, expressions, mouth movement and posture stay locked across the whole video — avoiding the identity breakdown and lip-sync slips of stitch-together clips. Stable across English, Spanish and Chinese.

The caveat: limits vary by plan

The full 30-minute cap applies on plans like Creator; shorter caps can apply for some uses. Generation time also rises with length and number of scenes.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.