Technical reports
Abstract grainy textured composition with overlapping geometric planes

Avatar 6: First On All 6 Criteria, and On Overall Score

Talking avatars that are more like you. From a single photo — and, optionally, a short recording as an identity reference — Avatar 6 reproduces how a person looks, speaks, and moves, up to 2K, with no length limit.

By HeyGen ResearchNov 17, 202617 min read
Read the technical report

Identity
#1 / 9
most like the real person among 9 commercial avatar products, from a single photo (#1 / 4 from the person's footage)
All 6 criteria
#1
in all 6 criteria rated by external raters, against commercial avatar products
Turing test
84.4%
indistinguishable from real footage side by side
vs Seedance 2.5
72.4%
of side-by-side answers with a preference favour Avatar 6, from the person's footage (74.1% vs Wan 3.0)
Unlimited talking avatar video generation

Any identity. Native 2K. Unlimited duration.

Generate talking videos from a single photo, optionally with a short recording as an identity reference, across diverse identities and visual styles, at up to 2K and with no fixed duration limit.

Any avatar
Each created from a single photo: AI-generated characters and real people.
Native 2K
Every frame generated at 2880×1620.
Unlimited duration
25–28 min per video here, each one continuous generation from one audio track, no loops. Length follows the audio.
Motion prompt control

Gestures on cue.

Write the action next to the words. Avatar 6 performs it right on cue, in the person's own style.

“Hello”thumbs up
“Hello”
thumbs up
“the counter on the left”point to the side
“the counter on the left”
point to the side
“always happy to chat”hand on heart
“always happy to chat”
hand on heart
“some water”wave
“some water”
wave

Prompt Friendly pharmacist. On ‘Hello', give a thumbs up. On ‘the counter on the left', point to the side with your index finger. On ‘always happy to chat', place your right hand on your heart. On ‘Take a seat', open both arms wide. On ‘some water', wave your right hand.

“Welcome”arms open, palms up
“Welcome”
arms open, palms up
“grandparents”thumbs up
“grandparents”
thumbs up
“front desk”wave
“front desk”
wave
“Thanks for trusting us”hand on heart
“Thanks for trusting us”
hand on heart

Prompt Warm clinic host. On ‘Welcome', open both arms wide, palms up. On ‘grandparents', give a thumbs up. On ‘front desk', wave your right hand. On ‘Thanks for trusting us', place your right hand on your heart.

“not the melody”wag index finger
“not the melody”
wag index finger
“hear it”cup an ear
“hear it”
cup an ear
“no, slower”palm to the camera
“no, slower”
palm to the camera
“the taste, the choices and the feeling”count to three
“the taste, the choices and the feeling”
count to three

Prompt Playful storyteller. On ‘not the melody', wag your index finger. On ‘hear it', cup a hand behind your ear. On ‘no, slower', raise a flat palm to the camera. On ‘rain on a window', flutter your fingers downward. On ‘the taste, the choices and the feeling', count to three on your fingers.

Multi-speaker dialogue and podcasts

Two speakers, one frame.

A two-minute podcast, host and guest, generated in one go.

Wide shot, host speaking
Wide shot, host speaking
Guest speaking
Guest speaking
Guest listening
Guest listening
Wide shot
Wide shot
Talking video editing

Change a word. Keep the recording.

Only the edited words are regenerated, in the person's own voice. Everything else stays as filmed.

7.67#1 / 5
lip sync on the new words (original recording: 6.25)
0.913#1 / 5
face similarity (next best 0.907)
0.763#1 / 5
voice similarity (next best 0.739)
48.3 dB#1 / 5
untouched footage (next best 29.3 dB)
OriginalAvatar 6 edit
0:00
Avatar 6unchangedunchanged
0 s2 s4 s6 s8 s
“honestly better than most restaurants” · + thumbs up
OriginalAvatar 6 edit
0:00
Avatar 6unchangedunchanged
0 s2 s4 s6 s8 s
“half of the members” · + two fingers
OriginalAvatar 6 edit
0:00
Avatar 6unchangedunchanged
0 s2 s4 s6 s8 s10 s
“use our headlamps” · + point upward
new wordsregenerated by Avatar 6rows: change from the original
Word edit: “made a pizza” → “made a lasagna”
OriginalMiniMax H3-maxWan 3.0Seedance 2.5HappyHorse EditAvatar 6
0:00
MiniMax H3-max
Wan 3.0
Seedance 2.5
HappyHorse Edit
Avatar 6outside the edit: identical to the original up to video encoding noiseunchanged
frame of the difference images belowframe shown below
0 s2 s4 s6 s8 s10 s
All videos follow the timing of the Avatar 6 edit
All six editing metrics, better outward
editing metrics
Full table
MethodUntouched
PSNR (dB) ↑
Frames
changed ↓
Face
similarity ↑
Voice
similarity ↑
Lip sync
LSE-C ↑
Lip sync
LSE-D ↓
MiniMax H3-max19.5100.0%0.9070.7177.257.75
Wan 3.025.8100.0%0.8350.7275.708.89
Seedance 2.529.399.7%0.8830.7393.4610.89
HappyHorse Edit22.8100.0%0.7510.5432.4710.36
Avatar 648.334.4%0.9130.7637.677.19

Outside the edit, nothing changes

One frame after the edit, compared with the original: the brighter, the more it changed.

Show the difference images
difference from the original
OriginalMiniMax H3-maxWan 3.0Seedance 2.5HappyHorse EditAvatar 6
Evaluation

Preferred over the strongest video models.

Against Seedance 2.5, Wan 3.0, our previous models and commercial avatar products. Also first on Avatar Arena.

Judged by the people in the videos

HeyGen employees compare videos of themselves and pick the one that looks most like them, which only they can tell. Below: online users vote between Avatar 6 and our previous models on the same request.

pref_general
Objective scores
radar_obj_models
radar_obj_platforms
Ratings by external raters: from a photo and from footage
acr_radar_photo
acr_radar_footage
Judged by strangers, side by side

External raters who have never met the person compare two videos for overall quality, lip sync and motion. Without knowing the person, likeness counts for less, so the margin over Seedance 2.5 and Wan 3.0 is smaller than above.

ccr_bars
Full rating tables
MethodOverallVisual
quality
IdentityLip
sync
Motion
naturalness
Gesture–
speech
Artifacts,
stability
Avatar 6
wins
From a single photo and speech
VEED Fabric 1.03.683.964.203.713.743.953.7069.6%
OmniHuman v23.323.703.973.443.413.483.3680.8%
Argil image avatar3.463.764.323.523.553.643.5880.9%
Creatify Aurora3.293.333.993.483.403.583.3284.3%
Akool Talking Photo3.193.524.113.283.333.353.4086.4%
Hedra Character 33.163.624.053.243.213.283.3388.2%
D-ID V2 Talks1.483.433.072.180.940.712.4297.9%
Colossyan NEO 1 Instant avatar1.593.403.481.761.190.782.5798.1%
Avatar 64.004.334.394.124.014.074.08–
From the person's own footage
Akool Talking Avatar2.933.703.892.782.852.723.2992.9%
D-ID V3 Express avatar2.853.743.832.602.762.713.2395.3%
Colossyan NEO 2 / NEO 12.683.593.652.562.692.483.2197.3%
  NEO 2, scripts 1–32.903.593.583.242.862.583.3596.2%
  NEO 1, scripts 4–52.353.603.741.552.432.343.0198.9%
Avatar 63.984.244.194.104.044.083.93–
Avatar 6 vsOverallVisual
quality
IdentityLip
sync
Motion
naturalness
Gesture–
speech
Artifacts,
stability
Avatar 6
wins
Photo input
Wan 3.0§−0.01+0.55+0.09+0.01−0.18−0.19+0.0450.1%
Seedance 2.5§+0.30+1.05+0.08+0.03+0.12+0.10−0.0159.6%
Kling 4.0+0.79+0.66+0.14+0.42+0.69+0.81+0.3073.5%
Synthesia Express-2 photo avatar+1.50+0.96+0.55+1.21+1.56+1.51+0.5887.0%
Footage input
Kling 4.0 reference-to-video§+0.17+0.73−0.08+0.04+0.01−0.02+0.0254.9%
Seedance 2.5 reference-to-video§+0.61+1.10+0.40+0.14+0.18+0.21+0.0372.4%
Wan 3.0 reference-to-video§+0.75+0.54+0.88+0.31+0.31+0.35+0.2974.1%
VEED Preset avatars‡+1.15+0.71+0.17+0.94+1.12+1.18+0.5887.6%
Creatify AI Avatar v1‡+1.31+0.84+0.25+1.12+1.30+1.37+0.7093.3%
Synthesia Personal video avatar†+1.68+1.64+0.79+1.38+1.51+1.56+1.1693.6%
Synthesia Stock avatars v3 / v1‡+1.28+0.92+0.24+1.01+1.26+1.37+0.6593.7%
  v3, four avatars+1.29+0.88+0.23+1.02+1.30+1.39+0.6292.8%
  v1, Cooper+1.26+1.05+0.28+1.00+1.12+1.27+0.7597.5%
Turing test

Comparison with Real Footage: Turing Test

Side by side with the real recording, raters pick Avatar 6 as the real one 42.2% of the time. 50% would mean they cannot tell at all, so Avatar 6 is 84.4% of the way there.

84.4%
indistinguishable from real footage side by side (Avatar V: 44.4%)
65.1%
of Avatar 6 clips judged real when shown alone (real recordings: 78.9%)
56.5%
still judged real with a HeyGen Voice clone
turing_design
Full results
Single clip, shown aloneJudged realAnswers
Real recording78.9%612
Avatar 6, person's own voice65.1%304
Avatar 6, HeyGen Voice56.5%308
Side by side, same sentence and audioReal pickedAnswers
Real vs Avatar 657.8%270
Avatar 6 picked as real42.2%270
Indistinguishability (2× the above)84.4%270