Video
HeyGen vs. Synthesia avatar models compared (2026)

Key takeaways
HeyGen's newest model, Avatar 6, launched on November 17, 2026. It animates a single photo or a short recording of you, and its research page lists exports up to 2K.
Synthesia's newest model, Express-3, launched on July 15, 2026 and is available on all plans. Synthesia's pages don't state a resolution, frame rate, or clip limit for it.
Both makers route gestures through the audio or script. HeyGen also takes written motion prompts, while Synthesia's pages describe gestures as script-aligned and don't mention a prompt field.
Synthesia's Express-1 is built from filmed actors with no gestures, and HeyGen's Avatar V learns movement from your own reference video. Those are the two clearest opposite approaches.
Most of the realism numbers come from each maker's own testing, so the documented specs are the safer basis for choosing.
Which avatar model is better, HeyGen's or Synthesia's?
Neither wins on every job, and the documents point to different strengths. HeyGen's models start from a photo or a clip of a specific person and aim to reproduce how that person moves, while Synthesia's latest models are built around stock and generated avatars that gesture to match a script.
Both makers keep several model generations live at once, so the first step is knowing which generation a given avatar uses. HeyGen's lineup is Avatar III, IV, V, and 6. Synthesia's is Express-1, Express-2, and Express-3, alongside older Classic and legacy avatars. This piece compares the proprietary models only. It leaves out templates, editors, and plan tiers except where a model is gated by one.
HeyGen's avatar models
HeyGen's developer model list names four engines. Avatar III is a photo-to-video pipeline that outputs up to 4K. Avatar IV is the default engine and is described as having the broadest coverage. Avatar V is labeled "Highest fidelity" and is opt-in per look. Avatar 6 is "Newest avatar engine. Up to 2K."
Avatar IV takes a photo or an existing avatar plus a script or an audio file. HeyGen's help center says it interprets vocal tone, rhythm, and emotion and not only mouth shapes, and its motion covers head tilts, natural pauses, and micro-expressions. On the API, Avatar IV accepts arbitrary images and has both a motion prompt and an expressiveness setting for photo avatars and images. HeyGen also says Avatar IV doesn't render native 4K through the API, so a 4K request is an upscale of a 1080p render.
Avatar V works differently. Its research page describes a model built on a diffusion transformer that takes one reference video and a driving audio signal, and it reports capturing "habitual micro-expressions, and gestural tendencies during speech" from that footage. The same page says voice cloning needs as little as 10 seconds of audio and that output can run to arbitrary length. HeyGen's help article adds that Avatar V supports videos up to 3 minutes when used in Video Agent. On the API, Avatar V supports motion prompts and a reference look from the same avatar group, but not the expressiveness setting.
Avatar 6 is the newest. HeyGen's technical report, dated November 17, 2026, says it works from a single photo plus an optional short recording as an identity reference. Gestures are cued by motion prompts written alongside the speech, such as a thumbs up, pointing, or counting on fingers, and the model supports dialogue between multiple speakers. Every frame is 2880 by 1620, exports go up to 2K, and the report says duration has no fixed limit, with examples running 25 to 28 minutes from one audio track. Creating an avatar from footage requires an on-camera consent statement with a one-time code and an identity match check. The report's realism figures, including 84.4% judged indistinguishable side by side, come from HeyGen's own evaluations.
Synthesia's avatar models
Synthesia names its generations Express-1, Express-2, and Express-3. Its help center describes Express-1 stock avatars as built from "professionally filmed footage of real actors." They adjust facial expression and tone of voice to the script, but they don't gesture. Custom Studio avatars also run on Express-1 and need filmed footage. Synthesia's avatar type guide lists them as a Starter-annual add-on or Enterprise option at $1,000 per avatar for self-submitted footage. That guide also lists Express-1 avatars as the ones compatible with the Synthesia API.
Express-2, announced on September 4, 2025, added full-body motion. Synthesia's research page describes a system with three models, including Express-Animate, a model for generating gestures driven purely by audio. It says Express-2 "generates 1080p videos of arbitrary duration at 30fps." The same page admits the unoptimized renderer was "slow and not practical for customer use," and says production generation takes around 8 minutes for a minute of 1080p video. Express-2 also paired with Express-Voice, which clones a voice from reference audio without fine-tuning. The avatar guide lists personal avatars made from a single uploaded photo as Express-2 based and available from the Starter plan up.
Express-3 is the current model. Synthesia's launch post, dated July 15, 2026, calls it "our most advanced avatar model for lip sync, natural body movement and gestures." It says videos can generate up to 2x faster, that Express-3 is available on all plans, and that it's rolling out in stages, starting with new stock avatars. Synthesia's update page says Express-3 replaces Express-2 across the platform and leaves Express-1 unchanged. Style Avatars, built on Express-3, can be made from a prompt and a preset, with a reference image option listed as coming soon.
How the models compare
The table below uses only what each maker's pages state. "Not stated" means the maker's pages don't say, and it shouldn't be read as a limit.
Model | Input | Output limits | Gestures | Status |
|---|---|---|---|---|
HeyGen Avatar IV | Photo or existing avatar, plus script or audio | 1080p on the API, 4K is upscaled | Head and facial motion, motion prompt, expressiveness setting | Default engine |
HeyGen Avatar V | One reference video plus audio | Arbitrary length, up to 3 minutes in Video Agent | From your own footage, motion prompt | Opt-in per look |
HeyGen Avatar 6 | Photo, optional short recording | Up to 2K, no fixed duration | Prompt-cued, multi-speaker | Launched November 17, 2026 |
Synthesia Express-1 | Filmed actors or studio footage | Not stated | None | Unchanged by Express-3 |
Synthesia Express-2 | Stock, or a single photo | 1080p, 30fps, arbitrary duration | Full-body, driven by audio | Being replaced by Express-3 |
Synthesia Express-3 | Stock, Style Avatars from prompt and preset | Not stated | Script-aligned body movement | Launched July 15, 2026 |
Two differences stand out. HeyGen documents motion prompts on its API for Avatar IV, V, and 6, which gives a developer a way to direct a gesture. Synthesia's pages describe gestures as following the script, and they don't mention a prompt field. And HeyGen's newest model is built around a specific person's likeness, while Synthesia's newest is introduced mostly through stock and generated avatars.
Languages don't separate the models in the documents. HeyGen's marketing for Avatar V says videos can star you in 177+ languages and dialects, and Synthesia's pages say 140+ languages. Neither maker breaks those counts down by model.
Which model to choose by use case
For a video of yourself or a named presenter from little footage, use HeyGen Avatar V or 6. Both are built around reproducing one person, and Avatar V needs a single reference video.
For a talking photo with direct control over a gesture, look at HeyGen Avatar IV or 6. Both accept photo input and a motion prompt on the API.
For long single takes, HeyGen Avatar 6 documents examples of 25 to 28 minutes, and Synthesia's Express-2 page documents arbitrary duration at 1080p. Avatar V is documented to 3 minutes in Video Agent.
For a presenter with no filming at all, Synthesia Express-3 stock avatars are available on every plan.
For automated pipelines through an API with a filmed avatar, Synthesia Express-1 fits, since its help center lists it as API compatible, as do HeyGen's engines through the developer API.
For full-body gestures from a script alone, consider Synthesia Express-2 or 3, whose pages describe gestures generated from audio or script.
If a job needs a hard spec the documents don't state, such as a 4K deliverable from Synthesia or a maximum clip length for Express-3, the answer isn't known from the published pages. Check the maker's current docs before committing to a schedule. Output artifacts like flicker and morphing are a separate problem, and our guide to fixing flicker and morphing in AI video covers it. For comparing avatar tools aimed at influencer-style content, see our roundup of AI UGC video generators, and for dubbing a presenter into other languages, the AI voice translators list. If you're deciding between starting from text or a picture, read text-to-video vs. image-to-video.
HeyGen vs. Synthesia avatar models FAQ
Which has the more realistic avatar, HeyGen or Synthesia?
No independent head-to-head is in the maker documents. HeyGen reports its Avatar 6 results from its own raters, and Synthesia's Express-3 pages make no independent realism claim. Treat the numbers as vendor claims.
Can I make an avatar from a single photo?
On HeyGen, yes. Avatar IV accepts photos, and Avatar 6 works from a single photo with an optional short recording. On Synthesia, the avatar guide lists personal avatars from one uploaded photo, built on Express-2, as available from the Starter plan up.
What are the video length limits?
They vary by model and product. HeyGen documents up to 3 minutes for Avatar V in Video Agent and no fixed duration for Avatar 6. Synthesia documents arbitrary duration for Express-2 and states no limit for Express-3.
Do I need my own voice?
Not necessarily. HeyGen's Avatar V page says cloning a voice needs as little as 10 seconds of audio, and Synthesia's Express-Voice clones a voice from reference audio.
Do both require consent to clone a real person?
Yes. HeyGen says creating a custom avatar requires explicit verification from the person represented, and Synthesia says it makes it impossible to clone a real person without their consent.
