Models

HeyGen's Avatar 6 edits 34.4% of frames where rivals rewrite nearly all

HeyGen's Avatar 6 regenerates only edited spans, changes 34.4% of frames per edit, and claims 84.4% indistinguishability. Independent tests ranked it first, on short clips only.

HeyGen Launches Avatar 6 with Script-Aware Gestures and 2K Video
HeyGen Launches Avatar 6 with Script-Aware Gestures and 2K VideoClickerHappy / pixabay
By James Calloway6 min read

HeyGen says its new Avatar 6 model changes just 34.4% of frames in a typical edit, while MiniMax H3-max, Wan 3.0, Seedance 2.5 and HappyHorse Edit rewrite between 99.7% and 100% of them. The company introduced the audio-driven avatar model in a technical report published November 17, 2026, and made it available in its web app, mobile app and API from mid-November 2026. Avatar 6 generates video of a specific person from a single photograph, with an optional short identity recording, and HeyGen reports native 2K output with no fixed duration limit.

The claim matters because avatar video is becoming a production tool for marketing and training content, and edit-localized regeneration decides whether a finished recording can be corrected or must be recreated. HeyGen's numbers come from its own 20-edit comparison, and the company has every incentive to frame them favorably. What the comparison does not say is how the edited spans themselves compare in quality.

What does Avatar 6 actually do?

Avatar 6 reproduces a person's appearance, manner of speaking and habits of motion together, according to HeyGen. Its in-house pipeline combines the video generator, a super-resolution stage that restores details such as skin texture and hair, a motion planner, and HeyGen Voice for speech in the speaker's own voice. Customers choose between Avatar 6 Quality and Avatar 6 Turbo, the latter using HeyGen's accelerated inference stack for lower latency.

The model's headline feature is script-linked gesture control. HeyGen says Avatar 6 uses a script's meaning to decide where pointing, counting or directional movement fits. Users can separately specify mood, facial expression or hand movement, and written motion prompts can attach an action to a particular phrase, such as waving during a greeting or counting on the fingers through a list. The motion planner times those actions to the relevant words while preserving the subject's characteristic movement between them.

Long-form output is the second pillar. HeyGen's demonstrations include continuous videos lasting 25 to 28 minutes, each generated from one audio track without loops, so the running time follows the speech. The same framework produces two speakers in one frame; one demonstration is a two-minute host-and-guest podcast that alternates wide shots with closer views of each person.

The recorded-video editor can replace words or add a gesture while regenerating only the edited span, synthesizing new speech in the person's voice. That preservation of the surrounding take is what makes correcting a finished presenter recording practical.

Takeaway: Avatar 6's pitch is control and repair, not just generation. But the 25-minute demos are HeyGen's own demonstrations, not independent results.

How real is the realism claim?

HeyGen's 84.4% indistinguishability score has a specific meaning that the number alone does not convey. External raters selected Avatar 6 as the real recording in 42.2% of 270 side-by-side answers, and the score doubles that selection rate because 50% would represent chance. The side-by-side test involved 20 external raters and 76 people.

In separate single-clip tests, 65.1% of answers judged Avatar 6 clips using the person's own voice to be real, against 78.9% for genuine recordings. That gap is the honest measure: raters still detect the synthetic clips a meaningful share of the time.

HeyGen's own external-rater comparisons were narrower against some general video models:

  • Among answers with a preference, Avatar 6 won 50.1% against Wan 3.0 from a photograph, described by HeyGen as a tie
  • 54.9% against Kling 4.0 from footage
  • 72.4% against Seedance 2.5 from footage

A tie against Wan 3.0 on photo-based generation is notable given Avatar 6 is purpose-built for avatars and Wan 3.0 is a general video model.

Takeaway: The indistinguishability headline is a derived figure, and the single-clip numbers show a detectable gap remains.

What did independent testing find?

New Measure's early Avatar Arena results ranked Avatar 6 first in both of its categories, with Elo-scale preference scores of 1,197 for personal avatars and 1,237 for stock presenters; each category averages 1,000. Creators compared two videos reading the same script with provider names hidden and video order randomized. The counted votes and edits ran from October 27, 2026, through November 16, 2026.

New Measure itself attaches significant caveats. Participants chose to use the arena, the typical script was 62 words, and both videos lasted 35 seconds or less in 97% of completed comparisons. The results do not establish performance on long videos or in languages with insufficient testing.

Preference also did not mean speed. HeyGen personal-avatar videos typically took about three minutes, versus about one minute for Runway GWM-1 Avatars. One in ten HeyGen personal-avatar videos took more than nine and a half minutes. The timings exclude avatar setup and speech preparation.

Takeaway: The independent ranking supports Avatar 6 on short clips, but the long-form demonstrations that distinguish it remain untested by third parties.

What are the API limits and consent rules?

The API documentation lists two model identifiers, heygen-avatar-6 and heygen-avatar-6-turbo, with 1080p as the default resolution. A Model API request combines a reference image or video with speech audio. The Application API starts from a HeyGen avatar look and a script with voice or audio.

The Application API supports Avatar 6 only for single-avatar requests. It excludes raw-image jobs, studio scenes, batch creation and scene editing. Omitting the engine selection continues to use Avatar IV, the previous generation.

Consent rules depend on the input type:

  • Before a footage-based HeyGen avatar can render, the person must record an on-camera consent statement containing a session-specific one-time code, and HeyGen verifies the consent speaker matches the training-footage subject
  • Photo references are available to all API workspaces with the person's permission
  • Realistic video references are enterprise-only
  • The documentation says video-reference consent is verified before submission, although some contracts can operate without that step after case-by-case review

That last clause deserves attention. A case-by-case path around consent verification is exactly the kind of exception that misuse scenarios exploit, and HeyGen does not describe who reviews those cases or by what criteria.

Takeaway: The consent architecture is layered, but the enterprise exception leaves a documented gap in the verification chain.

What is the so-what?

Avatar 6 gives production users more control over identity, motion and revision, with continuous output extending well beyond short presenter clips. The independent arena supports its short-clip appeal but leaves the long-form demonstrations untested. Its timings also show that a preferred avatar can still be slower to produce than the fastest rival, which matters for any workflow with turnaround deadlines.

Source: unite.ai

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

722 articles

Related articles

  1. Google Cloud ships Gemini Omni 1.2 Flash at $0.07 per second
  2. Google Upgrades Veo 3.1 With 4K Upscaling and Native Vertical Video
  3. Ideogram 4.5 Targets Surgical Image Editing at 0.8 Cents a Picture
  4. Google Launches Gemini Omni Flash, a Video Model You Edit by Talking
  5. Suno's Visuals turns AI songs into AI video, up to 15 seconds per clip

« Previous articleNext article »