Stock avatar, digital twin, or generated character
Library avatar, trained twin, or generated character: how each one scores on setup, consent, brand ownership, and the way it ages out from under you.
The three ways to put a face on camera without booking a shoot are not three price points on one product. They are three different assets with three different owners, three different failure modes, and three different answers to the question a lawyer will eventually ask, which is: who gave you permission, and permission to do what.
Most teams pick on setup speed, discover the rights problem eighteen months later, and find that the cheapest option to start was the one they cannot keep.
The three options, side by side
A library avatar is a face the platform licensed and made available to everyone. On Versely that is the built-in avatar library — static preview images used as reference input for avatar-driven generations — alongside the premade avatar models and the roster of HeyGen digital-twin avatars available to the lipsync tooling. You pick a name from a list and you are recording within a minute.
A digital twin is a specific real person, recorded once and turned into a reusable presenter. This is the HeyGen Avatar V5 digital twin path and its V3 predecessor: consented footage of a real human goes in, a presenter that can say anything comes out.
A generated character is a face that does not correspond to anyone. You build it from a prompt and a reference set, then hold it steady across generations with reference images or a trained adapter.
| Library avatar | Digital twin | Generated character | |
|---|---|---|---|
| Setup before the first video | None. Pick from a list | A consented recording session, then training turnaround | Prompt and reference set, iterated until the face is right |
| Whose likeness | The platform's licensed performer | A real, identifiable person | Nobody's, if the build was clean |
| Rights you must hold | Whatever the platform's terms grant | Full grant, including derivative and term | Your own prompt and reference chain |
| Brand distinctiveness | None. Competitors use the same faces | High, and tied to one person | Highest, and transferable |
| How it ages | Vendor deprecates or replaces it | Ages with the person, leaves when they leave | Does not age, which is its own tell |
| Consistency mechanism | Fixed by construction | Fixed by construction | Re-supplied every time, from references or an adapter |
The rights question to settle before anyone records anything
There is one question that decides whether the digital twin path is available to you, and it has to be answered before the shoot day, not after. Most teams answer three quarters of it and think they are done.
Does your talent release cover training a model, or only using the recording?
Those are different grants. A standard talent release written any time before generative video was normal grants the right to use, edit and distribute the footage that was captured. It does not necessarily grant the right to derive a system from that footage that generates performances the person never gave. A twin trained on footage you hold a display licence for is an asset you cannot safely publish, and no amount of downstream care fixes it.
Four sub-questions, all of which need a written answer:
- Scope. Organic only, or paid too? Usage rights default to the narrowest reading when the contract is silent — posting to your own account and spending media budget behind it are different grants, and silence favours the person, not the brand.
- Term. How long after they leave can you keep publishing? Does the grant survive termination of employment, and does "in perpetuity" actually appear, or did everyone assume it?
- Revocation. Can they withdraw consent, and if they do, what happens to assets already live versus new generations? These usually need different answers, and a contract that treats them as one clause will be read against you.
- Derivative. The one above. Does the grant explicitly cover training, fine-tuning and synthesis, by name?
A library avatar sidesteps all four, because the platform did the clearing. That is genuinely the strongest argument for starting there. But note what you are buying: a face that is simultaneously fronting other brands' ads, in a library the vendor versions on its own schedule. The consent problem is solved and the distinctiveness problem is created.
A generated character sidesteps all four differently, by never involving a person. That holds only if the build is clean — a character generated from a reference photo of a real individual is a likeness question wearing a different hat, and the platform-side likeness and consent regimes now emerging around synthetic presenters, like TikTok's Symphony avatar licensing, are written to catch exactly that.
How each one ages, which nobody prices in
Setup cost is a one-time number and it gets all the attention. Ageing is a recurring cost and it gets none.
Library avatars age by vendor decision. The library is versioned by someone else. An avatar you built a year of content around can be replaced, retired, or updated into a slightly different face, and your back catalogue no longer matches your new uploads. You have no vote and usually no notice.
Digital twins age with the person, twice. First biologically: haircut, glasses, weight, a beard. A twin trained in January is visibly a year out of date by the following January, and viewers who see the person in real life notice before you do. Second, organisationally: the twin leaves when the person leaves, and if the derivative clause was vague, it leaves faster than that. There is also a model-version axis — a V3 twin and a V5 twin are not the same asset, and moving between them is a re-record, not an upgrade button.
Generated characters do not age at all, which is the quiet problem. A face that has been identical for three years across two hundred videos reads as synthetic even to viewers who cannot articulate why. The countermeasure is to deliberately version the character: wardrobe changes, a seasonal look, occasional variation in the reference set. That is work, but it is work you control, which is more than the other two offer.
Consistency is also the generated character's standing tax. Nothing carries between generations by default, so the same face has to be re-supplied every time from reference images or a trained adapter that makes the subject available from a prompt keyword. That is the difference between a character and a description of a character, and it is why a generated presenter takes more work per video to keep on-model than a twin or a library face, which are fixed by construction.
A decision rule
Answer in this order and stop at the first yes.
- Do you need a specific real person's authority on camera? Founder-led content, a named expert, a spokesperson viewers already recognise. Digital twin, and settle the four rights questions before booking. Nothing else substitutes, because the value is the identity.
- Is this content you expect to still be running in two years? Generated character. The ownership and transferability are worth the consistency tax, and it is the only option where nobody can end your access to your own presenter.
- Do you need volume this week and the face is interchangeable? Library avatar. Product explainers, internal comms, ad variants where the presenter is a delivery mechanism rather than a brand asset. Accept the distinctiveness cost knowingly rather than discovering it in a competitor's ad.
- None of the above? Skip the face entirely. A voiceover over b-roll clears every rights question in this post and is frequently the better creative decision anyway.
Whichever you land on, the disclosure obligation is the same and it does not vary by which of the three you chose. A synthetic presenter is a synthetic presenter whether the underlying face is licensed, cloned or invented, and synthetic media disclosure rules attach to the output, not to the provenance of the input. Build the label into the template rather than into a checklist someone remembers.
FAQ
Can I train a digital twin on footage I already have?
Only if the release covering that footage grants derivative and training rights explicitly. Archive footage is the most common place this goes wrong, because it was shot under a release written for a world where "use the video" meant playing it. Re-papering with the person is cheap if they are still reachable and impossible if they are not, which is a good reason to audit your archive's releases before you plan a campaign around them.
Is a generated character safe from likeness claims by default?
Not automatically. It is safe if the character was built from prompts and synthetic references with no real individual in the chain. It is not safe if you seeded it from a photograph of someone, however loosely, or if the result is recognisably a specific person. Keep the reference set and the prompt history — provenance you can produce is worth considerably more than provenance you can assert.
How many reference images does a generated character actually need?
Fewer than people think, and more varied than people supply. Models publish a ceiling on how many references they will hold at once, and beyond it extra images are dropped rather than averaged in. Within the ceiling, variety helps more than volume: three angles of one face beats six near-identical frames, because near-identical references teach the model the lighting and the room as well as the person.
Can I switch from a library avatar to a twin later without redoing everything?
The content already published stays as it is, and that is usually fine. What does not transfer is audience recognition — viewers who learned your brand through one face experience the switch as a personnel change, and the back catalogue now features someone who no longer works there. If a twin is the eventual destination, going there early is cheaper than going there at scale, and avatar-driven generation works the same way regardless of which of the three you feed it.