Video modelActive
Avatar V
HeyGen
- Released
- April 15, 2026
- Data date
- October 7, 2026
- Model type
Avatar V separates a recorded performance from the chosen appearance. A 15-second clip supplies expressions and movement while outfits, settings, and other looks can change; identity is designed to remain stable across camera angles and longer videos.
Specifications and access
| Specification | Value and source |
|---|---|
| Model type | Proprietary modelSource |
| Introduction | HeyGen introduced Avatar V on April 15, 2026Source |
| Identity input | A 15-second reference clipSource |
| Identity consistency | Maintains facial identity across camera angles, looks, and longer videosSource |
| Access | HeyGen web product; the announcement links to try itSource |
| Input modalities | 15-second video reference clipSource |
| Output modalities | VideoSource |
Page 1 of 2
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Pairwise human preference vs Kling O3 Pro (Show measurement, test conditions, and source)
- Source value
- 69.6
- Score
- 69.6
- Metric
- cases where majority vote preferred Avatar V
- Unit
- %
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 69
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen pairwise human preference test on identical inputs, with Avatar V and one named competitor shown side by side. Three annotators judged each pair and majority vote determined the winner; case counts vary because some commercial APIs rejected likeness inputs. Competitor: Kling O3 Pro.
- Context
- Vendor-run head-to-head evaluation. API refusals changed the number of cases by competitor; this is not a broad independent leaderboard.
Pairwise human preference vs Seedance 2.0 (Show measurement, test conditions, and source)
- Source value
- 68.9
- Score
- 68.9
- Metric
- cases where majority vote preferred Avatar V
- Unit
- %
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 45
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen pairwise human preference test on identical inputs, with Avatar V and one named competitor shown side by side. Three annotators judged each pair and majority vote determined the winner; case counts vary because some commercial APIs rejected likeness inputs. Competitor: Seedance 2.0.
- Context
- Vendor-run head-to-head evaluation. API refusals changed the number of cases by competitor; this is not a broad independent leaderboard.
Pairwise human preference vs Veo 3.1 (Show measurement, test conditions, and source)
- Source value
- 72.5
- Score
- 72.5
- Metric
- cases where majority vote preferred Avatar V
- Unit
- %
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 40
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen pairwise human preference test on identical inputs, with Avatar V and one named competitor shown side by side. Three annotators judged each pair and majority vote determined the winner; case counts vary because some commercial APIs rejected likeness inputs. Competitor: Veo 3.1.
- Context
- Vendor-run head-to-head evaluation. API refusals changed the number of cases by competitor; this is not a broad independent leaderboard.
Pairwise human preference vs OmniHuman 1.5 (Show measurement, test conditions, and source)
- Source value
- 85.7
- Score
- 85.7
- Metric
- cases where majority vote preferred Avatar V
- Unit
- %
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 70
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen pairwise human preference test on identical inputs, with Avatar V and one named competitor shown side by side. Three annotators judged each pair and majority vote determined the winner; case counts vary because some commercial APIs rejected likeness inputs. Competitor: OmniHuman 1.5.
- Context
- Vendor-run head-to-head evaluation. API refusals changed the number of cases by competitor; this is not a broad independent leaderboard.
Avatar Turing Test: real/fake identification accuracy (Show measurement, test conditions, and source)
- Source value
- 77.8
- Score
- 77.8
- Metric
- real/fake identification accuracy; lower indicates harder detection
- Unit
- %
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Lower is better
- Sample
- 54
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen Avatar Turing Test: three annotators each evaluated 18 pairs of Avatar V outputs and ground-truth videos, for 54 judgments. Lower real/fake identification accuracy indicates less successful identification of generated output.
- Context
- Provider-run test with only 18 pairs. The stated 50% chance level is a baseline, not an Avatar V result.
Avatar Turing Test: fooled rate (Show measurement, test conditions, and source)
- Source value
- 22.2
- Score
- 22.2
- Metric
- judgments where Avatar V output was mistaken as real
- Unit
- %
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 54
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen Avatar Turing Test: three annotators each evaluated 18 pairs of Avatar V outputs and ground-truth videos, for 54 judgments. Lower real/fake identification accuracy indicates less successful identification of generated output.
- Context
- Provider-run test with only 18 pairs.
Avatar Turing Test: cases fooling at least one annotator (Show measurement, test conditions, and source)
- Source value
- 61.1
- Score
- 61.1
- Metric
- pairs where at least one annotator mistook Avatar V for real
- Unit
- %
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 18
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen Avatar Turing Test: three annotators each evaluated 18 pairs of Avatar V outputs and ground-truth videos, for 54 judgments. Lower real/fake identification accuracy indicates less successful identification of generated output.
- Context
- This is a pair-level count, separate from the 54 individual judgments.
Cross-scene benchmark: SyncNet Sync-C (Show measurement, test conditions, and source)
- Source value
- 8.97
- Score
- 8.97
- Metric
- video-level synchronization confidence
- Unit
- No unit provided
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 36
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen cross-scene benchmark; objective table uses the 36 matched cases where all five systems produced valid outputs, drawn from a 70-case set. This SyncNet metric is reported at video level.
- Context
- HeyGen is the evaluator and model provider. The reported cohort and setup are specific to the vendor report.
Cross-scene benchmark: SyncNet Sync-D (Show measurement, test conditions, and source)
- Source value
- 6.75
- Score
- 6.75
- Metric
- video-level minimum distance
- Unit
- No unit provided
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Lower is better
- Sample
- 36
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen cross-scene benchmark; objective table uses the 36 matched cases where all five systems produced valid outputs, drawn from a 70-case set. This SyncNet metric is reported at video level.
- Context
- HeyGen is the evaluator and model provider. The reported cohort and setup are specific to the vendor report.
Cross-scene benchmark: Face Similarity (Show measurement, test conditions, and source)
- Source value
- 0.84
- Score
- 0.84
- Metric
- ArcFace embedding cosine similarity
- Unit
- ratio
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 36
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen cross-scene benchmark; objective table uses the 36 matched cases where all five systems produced valid outputs, drawn from a 70-case set. This SyncNet metric is reported at video level. Face and perceptual frame scores are sampled at 2 fps and computed as 10% trimmed means.
- Context
- HeyGen is the evaluator and model provider. The reported cohort and setup are specific to the vendor report.
Cross-scene benchmark: Q-Align IQA (Show measurement, test conditions, and source)
- Source value
- 4.85
- Score
- 4.85
- Metric
- Q-Align score calibrated to human mean opinion scores
- Unit
- rating
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Sample
- 36
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen cross-scene benchmark; objective table uses the 36 matched cases where all five systems produced valid outputs, drawn from a 70-case set. This SyncNet metric is reported at video level. Face and perceptual frame scores are sampled at 2 fps and computed as 10% trimmed means.
- Context
- HeyGen is the evaluator and model provider. The reported cohort and setup are specific to the vendor report.
Cross-scene human MOS: Identity consistency (Show measurement, test conditions, and source)
- Source value
- 4.98
- Score
- 4.98
- Metric
- mean opinion score on a 1-5 Likert scale
- Unit
- rating
- Benchmark version
- HeyGen Avatar V technical report, 2026-04-08
- Category
- video-avatar
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- HeyGen Research Avatar V technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HeyGen human MOS evaluation on generated Avatar V videos, five-point Likert scale, 1 = very poor and 5 = excellent. Each video was rated by at least two trained annotators blinded to model identity, in randomized order; report gives the arithmetic mean. The report does not state the exact number of MOS-rated videos.
- Context
- The provider report does not publish the exact MOS item count; results are vendor-reported.
Page 1 of 2
Sources and data date
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 7, 2026. 8 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not found in the inspected sources | independent Avatar V benchmark replication Original wording The located numeric results are from HeyGen Research, the provider. No independent evaluator result for the exact Avatar V system was located in the checked sources. |
| Additional source | HeyGen Avatar V model details (retrieved October 7, 2026) · HeyGen Avatar V announcement · Editorial description reviewed October 7, 2026 |
| Additional source | HeyGen dated Avatar V introduction (retrieved October 7, 2026) |
| Additional source | HeyGen Research Avatar V technical report (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |