Two routes. Generate an actor from a text description, which depicts nobody real and needs no consent. Or upload a face, which depicts whoever is in the photograph — including you, which is the intended case. Either way the actor saves to your library with a handle you call in every prompt, across every tool. Both are free to cast; only real human creators add a fee.
Picking from a library is faster. Making your own is the right answer when the character has to be specific, has to recur, and has to be yours.
Where do you create one?
Inside the actor picker, as a tile rather than a separate flow.
Open the picker and you'll see the type filters — All, AI actors, Human creators, My actors — alongside two tiles: Create actor and Faceless. So making one doesn't mean leaving what you're doing; it's an option at the point of casting.
Anything you create lands in My actors, which is your own library rather than the platform's.
What are the two routes?
They produce the same kind of asset and sit in completely different rights positions.
| Generate actor | Upload a face | |
|---|---|---|
| What you provide | A handle, a persona description, and a few fields | A handle, a photo, a voice, and three ticked attestations |
| Who's depicted | Nobody real | Whoever is in the photograph |
| What the product asks | Nothing beyond the description | That you own it, that it doesn't depict a real identifiable person, and that lookalikes may be rejected |
| Best for | Any character who isn't a specific real person | Read the attestations before deciding |
| Cost to cast | Free | Free |
Source: Hexcoded actor picker, Talking Actors documentation and render costs page, checked September 2026. Legal requirements vary by jurisdiction.
The two routes differ in what the product asks of you, not in what it produces.
What do the fields actually do?
Four of them, plus the description, and they're doing different jobs.
The handle is the identifier you'll call in prompts. It's the most consequential field and the one people rush — you'll type it hundreds of times, so make it short and unambiguous.
The persona description is where the specificity goes. Age range, build, hair, wardrobe, lighting sensibility. This is the actual casting.
Gender, age, primary language and vibe set the persona's frame. They constrain the generation rather than describing it, which means a detailed description plus contradictory fields will fight each other.
The description is where to spend your effort. The fields are where to be consistent with it.
How specific should the description be?
Specific on structure, and less worried about surface.
Face shape and build read at phone size. Hair colour and wardrobe don't, at least not as identity — they read as styling. So if a character has to be distinguishable from three others in a thirty-second episode, the structural features are what carry it.
The other thing worth naming is one unmistakable marker: a scar, glasses, a distinctive hairline. Something a viewer could put into words. That does more for recognisability than another paragraph of description.
Where this falls short. This is practice guidance rather than a measured finding. What's certain is that a description you write once and call by handle stays fixed, while a description you retype varies — and that's the mechanism that matters more than the wording.
What does uploading a photo involve?
Four inputs and three attestations, and the button won't unlock without all of them.
The photo. JPEG or PNG, minimum 300×300 pixels, with portrait 9:16 working best. The dialog shows the file and lets you replace it.
A handle, same rules as the generated route.
A voice, selected in the same dialog.
And three attestations, which is where this route stops being a formality.
What do the attestations ask?
Three checkboxes, and the Create actor button stays disabled until all three are ticked:
That you own the image, or that it's AI-generated by you or your brand.
That it does not depict a real, identifiable person.
That you understand images which look like real people may be rejected.
Read the second one carefully, because it isn't a consent checkbox. It doesn't ask whether you have permission — it asks you to confirm no permission is needed.
That's a stricter position than a consent tickbox, and it changes what going wrong looks like. Uploading a colleague's photograph without asking isn't a step you skip. It's an attestation you make falsely, on a gated control, having read a sentence that says the opposite of what you're doing.
It doesn't ask whether you have permission. It asks you to confirm there's nobody to ask.
Where this falls short. An attestation is a statement, not a verification. Nothing examines the file to see whose face is in it — the gate is on the claim rather than the image. So the obligation is still yours; what's changed is that you've put it in writing.
What happens after you create it?
It saves to your library with a handle, and that's the whole point.
Every actor you create takes an @handle — lowercase, digits, dashes or underscores, unique to your workspace. You call that handle in any prompt, in any tool. The actor you make for episode one is the actor you call in episode seventy, and it's the same library whether you're in Creative Studio, Talking Actors, Make a Custom Video or URL → Ad.
There's also an optional description field on the saved reference, up to 200 characters. Use it to record what the character is and what created it — a text description or a photograph, and whose.
That second part matters more than it looks. It's five seconds at the time, and it's the only evidence that a character you're calling synthetic genuinely has no real person inside it.
Create once, call the handle forever. Re-describing a character is what makes the face drift.
What should you test before writing around it?
Three angles, and they break in a predictable order.
Profile. The most common early failure and the one that ruins a conversation scene.
Backlit. Identity thins when the face is edge-lit rather than front-lit.
Mid-turn. A head in motion is the hardest thing to hold, and it's unavoidable in any scene with movement.
A face that only works front-on isn't cast, it's a headshot. Finding that out after six episodes are written is the expensive version of finding it out now.
When should you pick from the library instead?
Three cases.
When the face doesn't matter. Background, ensemble, anyone whose specifics are irrelevant. Thousands of AI actors are ready to cast, and creating one is work with no return.
When you need a real, verified person. That's a human creator — liveness check and face-match passed, consent on record, and they earn on every render. Creating your own doesn't get you that, and no amount of description substitutes for it.
When you're exploring. Picking is faster than creating, and seeing what's available often tells you what you actually want.
What does it cost?
Nothing extra, and that's worth saying plainly because it's not obvious.
Casting a real human creator adds a small flat fee. AI actors and an actor you made from your own photo don't — they're free to cast. So creating your own has no per-render premium at all.
What you still pay is the generation itself, driven by clip length and resolution tier, and the cost shows before you render.
Worth knowing that real human creators also need a plan above the entry tier, whereas creating your own doesn't.
Create the actor before the script is locked
Cast first, write around what you have. Writing a character and then discovering the face can't hold a profile is the wrong order.
Choose the handle carefully
Short, unambiguous, and something you'll still recognise at episode forty. You'll type it more than any other word in the project.
If you're uploading, read the attestations before ticking them
Three lines on a gated control, and the second one says something specific about whose face you're uploading. They take ten seconds to read.
Put the specificity in the description, not the fields
Structure over surface — face shape and build read at phone size, hair colour doesn't. Add one unmistakable marker a viewer could name.
Test profile, backlit and mid-turn
Three generations, before any script depends on the face. This is the cheapest quality check available.
Fill in the 200-character note
What the character is and what created it. Five seconds now, unrecoverable later, and it's the only record that a synthetic character is genuinely synthetic.
Current as of the publication date. Actor creation fields, library capabilities and plan features change. Hexcoded's terms and actor licence are governed by the laws of India, with jurisdiction in Gurugram, Haryana. Nothing here is legal advice.
- Two routes: generate from a description, or upload a face. The difference is rights, not quality
- A description depicts nobody real. A photograph depicts whoever is in it, including you
- Uploading a photo requires three attestations, and the button stays disabled until all three are ticked
- One of them asks you to confirm the image doesn't depict a real identifiable person. Read it before you tick it
- An attestation isn't a verification. Nothing examines the file to see whose face is in it
- Create actor sits inside the picker as a tile. You don't leave what you're doing
- Put the specificity in the description. The persona fields constrain rather than describe, and contradicting them fights the generation
- Structure over surface. Face shape and build read at phone size; hair colour reads as styling
- Test profile, backlit and mid-turn before any script depends on the face
- Creating your own costs nothing extra to cast. Only real human creators add a fee
Two routes, both from the Create actor tile in the actor picker. Generate an actor builds a persona from a text description plus a handle and a few fields. Upload a face takes a photo reference or an existing avatar. The first depicts nobody real; the second depicts whoever is in the photograph.
Rights, not quality. A generated persona depicts nobody, so no consent is needed. An uploaded face depicts a real person, which needs their documented consent — and there's no consent step in the product, so that obligation is yours.
That you own the image or generated it, that it does not depict a real identifiable person, and that you understand images resembling real people may be rejected. The Create actor button stays disabled until all three are ticked — so the product asks you to confirm no consent is needed rather than that you have it.
The second attestation asks you to confirm the image doesn't depict a real, identifiable person. Someone else's photograph does. Nothing examines the file, so the gate is on your claim rather than the image — which means uploading a colleague's photo means attesting to something untrue rather than skipping a step.
Specific on structure — face shape, build — and less worried about surface, since hair colour and wardrobe read as styling rather than identity at phone size. One unmistakable marker a viewer could name does more than another paragraph.
No. Casting a real human creator adds a small flat fee, while AI actors and an actor you made from your own photo are free to cast. You still pay for the generation itself, driven by clip length and resolution tier.
Profile, backlit and mid-turn. Those break first and in that order. A face that only works front-on isn't cast — and discovering that after six episodes are written is far more expensive than three test generations now.
Cast it once, call it forever
Create an actor from a description or a photo, give it a handle, and call that handle in every prompt across every tool. Free to cast, and it saves to your own library.
See the actor libraryMore on characters, consent and reusable elements in Elements.