Three sources in the picker: AI actors from the platform, human creators who are real licensed people, and my actors — your own library. You create an actor either from a text description or by uploading a face, and it saves to your library with a handle you call in any prompt. The tier decides your rights position. The handle decides whether the face survives the series.
Casting is a discipline in conventional production and it disappears in most AI workflows, which treat "pick a face" as one decision made once. It's two decisions — which tier, and whether you name it — and only one of them is about the face.
What are the three tiers?
The picker offers three sources, plus two options that aren't a face from the library at all.
AI actors are platform personas — thousands ready to cast, no real person involved.
Human creators are real people who submitted their own likeness under a licence, passed a liveness check and a face-match, and earn every time a paying brand generates with them. Every card in the picker carries a "Human" badge and a speaker icon, so you can audition the voice before you cast.
My actors is your own library: anything you've generated from a description or uploaded from a photograph.
Alongside those sit Create actor, which makes a new one without leaving the flow, and Faceless, which renders voiceover and product visuals with no presenter — the right answer more often than people expect.
The tier decides your rights position. The handle decides your continuity. They're separate decisions.
That second half is what most casting advice misses. Whichever tier you pick, a character re-described in words comes back slightly different every generation — each pass is independent and the model has no memory of the last one. Naming the actor is what removes the re-description.
Which tier for which job?
| Tier | What it is | Rights position | Best for |
|---|---|---|---|
| AI actors | Platform personas, thousands ready | No real person depicted, so no likeness obligation | Background, ensemble, anything where the face isn't the point |
| Human creators | Real people, licensed. Liveness check and face-match passed, consent-verified receipt shown at casting | Licensed. Digital use only, light edits only | Client work, anything a legal reviewer signs off on. Needs Starter or above |
| My actors | Your own — generated from a description, or uploaded from a photograph | Depends entirely on the source. A description carries nothing; a photograph carries whoever is in it | Recurring leads, and any character you'll use across a series |
| Faceless | No on-screen presenter | Nothing | When the product or the voiceover should carry the piece |
Source: Hexcoded actor library, Terms of Service and pricing page, checked September 2026. Legal requirements vary by jurisdiction.
Where this falls short. Most productions are mixed, and that's correct rather than a compromise. A human creator delivering to camera with AI actors filling the room behind them is normal. Picking one tier for everything is what makes casting expensive.
How do you actually create an actor?
Two methods, and the difference between them is a rights difference rather than a quality one.
Generate actor — from a description
You write a handle, describe the persona, and set a few fields including gender, age, primary language and vibe. The description is where the specificity goes — age range, build, hair, wardrobe, lighting.
Nobody real is depicted, so there's no consent question. This is the clean case, and it's the one most narrative work should default to.
Upload a face — from a photograph
You supply a photo reference or an existing digital avatar, and it becomes an actor in your library.
Whoever is in that photograph is now depicted in your work. If it's you, you're the consenting party and the paperwork is trivial. If it's anyone else — a colleague, a founder who hasn't been asked, a friend who agreed verbally — it's a likeness of that person and needs their documented consent.
What makes an actor persist across a series?
The handle. This is the single most useful thing in this post.
Every actor you create saves to your actor library with an @handle — lowercase, digits, dashes or underscores, unique to your workspace. You then call that handle in any prompt, in any tool. The actor built 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 an optional description field on each saved reference, up to 200 characters. Use it to record what the character is and what created it — it's the cheapest cast log you'll get, and exactly what you'd want if a resemblance question ever arose.
The same system covers places and looks. Characters, elements, styles and locations are all reference types, so the apartment your series lives in can be @apartment-kitchen rather than a paragraph you retype and accidentally vary.
Create the actor early, once
Generate from a description or upload a face, before the script is locked. Give it a handle you'll remember. This is the only time you describe the character.
Test it on the hard angles before you write around it
Profile, backlit, mid-turn. Those break first. A face that only works front-on isn't cast, it's a headshot — and finding that out after six episodes are written is the expensive version.
Use the handle, never a re-description
@maya in every prompt. Re-describing the character is a self-inflicted source of drift, and the most common cause of a face changing across a series.
Fill in the description field
Up to 200 characters recording what the character is and what created it. Five seconds now, unrecoverable later.
How many actors can appear in one scene?
More than you'd want, and fewer than the number suggests.
In Creative Studio's video engine, up to 30 actor references can sit in a single generation, inside an overall pool of 50 assets — 30 images, 30 actors, 10 videos and 10 audio, capped at 50 combined. That ceiling varies by model: Seedance 2.5 allows 50 total, Wan 2.7 allows 8.
The practical limit is lower than any of those numbers. Two characters interacting in one frame produce identity blurring on every platform currently available, so a scene with two of your cast in it is shot as over-shoulders and clean singles rather than a true two-shot. Each of those has only one identity to hold.
Where this falls short. The reference ceiling tells you what the model will accept, not what it will render well. Attaching thirty actors to one generation is technically possible and almost certainly a bad idea.
How do you cast an ensemble without them blending together?
The problem nobody warns you about: generate four supporting characters from similar descriptions and you get four variations on one face. Same age range, same bone structure, same lighting sensibility. On a phone screen at thirty seconds an episode, viewers can't tell them apart.
Conventional casting solves this without thinking about it, because you're choosing between real people who are already different. Generated casting has to solve it deliberately.
- Vary the structural features, not the surface ones. Face shape and build read at phone size; hair colour and wardrobe don't
- Give each character one unmistakable marker — a scar, glasses, a distinctive hairline. Something a viewer could name
- Check them side by side before you generate anything. Four faces in one frame at thumbnail size is the test
- Assign a wardrobe palette per character and keep it fixed. Colour is the fastest identifier when the face is small
- Never cast two similar characters into scenes together. The model degrades both, and the viewer conflates them anyway
When do you need a human creator?
When someone speaks to camera in client work, and when a legal reviewer has to sign off.
A human creator is a real person who agreed to this specific use. They passed a liveness check and a face-match at onboarding, accepted a licence with a timestamped record, and their likeness couldn't be published without clearing both. When you cast, you're shown a consent-verified receipt. That's a documented position rather than a convincing image, and the difference shows up at approval stage rather than in the edit.
What the licence permits is narrower than "commercial use" suggests. It's digital only — paid ads and organic posts on online platforms, plus your own websites and apps. Television, print and outdoor need a separate written agreement. And light edits only: trimming, cropping, captions, music and end-cards are fine, but you may not re-generate, morph or re-voice a face outside the platform.
One more thing worth planning for. A human creator can stop new use of their likeness at any time. Content already generated and delivered stays valid — but new generation stops, which for a long-running series means your lead could become uncastable mid-arc while the episodes you've made remain perfectly usable.
Where this falls short. You're choosing from a library rather than casting the exact face in your head, it costs more because a person is being paid on every render, and it needs Starter or above. For a background character or a personal project, that's the wrong trade.
What does casting get wrong most often?
Four mistakes, in rough order of how expensive they are to undo.
Not naming the actor. Re-describing a character in each prompt is the single most common cause of a face changing across a series, and it's entirely avoidable.
Casting from the hero shot. A face chosen from one flattering front-on generation will fail the first time you cut to a profile, and by then the episode is written around it.
Uploading someone else's photograph without asking. The tool won't stop you. The consent obligation is still yours.
Reusing a background face as a named character. It happens because the face was convenient. Viewers notice, and you now need continuity for a character you never planned to track.
Current as of the publication date. Actor library capabilities, reference ceilings, plan features and licence terms change. Hexcoded's terms and actor licence are governed by the laws of India, with jurisdiction in Gurugram, Haryana. Nothing here is legal advice.
- Three tiers: AI actors, human creators, and your own library. Plus faceless, which is right more often than people expect
- The tier decides your rights position. The handle decides your continuity. Separate decisions
- Create the actor once and call it by handle. Re-describing a character is what makes the face drift
- Generate actor from a description carries no likeness obligation. Upload a face carries whoever is in the photograph
- There's no consent step on upload. That makes consent your discipline, not the tool's
- Test profile, backlit and mid-turn before you write around a face. Front-on tells you nothing
- Vary face shape and build across an ensemble. Hair and wardrobe don't read at phone size
- Human creators need Starter or above, and they can stop new use at any time. Delivered episodes stay valid
By tier and by handle. Choose the source — platform AI actors, real human creators, or your own library — based on the rights position the job needs. Then save the actor with a handle and call that handle in every prompt, which is what keeps the face consistent.
Create the actor once, save it to your actor library with a handle, then call that handle in every prompt. The reference persists across generations and across all four tools, so you never re-describe the character — and re-describing is what causes drift.
Two ways. Generate actor builds a persona from a text description plus a few fields including gender, age, primary language and vibe. Upload a face takes a photo reference or an existing avatar. The first depicts nobody real; the second depicts whoever is in the photograph.
Only with their documented consent. There's no consent step on upload, so nothing in the product prevents it — which makes the consent your responsibility. A likeness built from someone's photograph depicts that person.
Up to 30 actor references in Creative Studio's video engine, inside an overall pool of 50 assets. That ceiling varies by model — Seedance 2.5 allows 50 total, Wan 2.7 allows 8. The practical limit is much lower, because two characters sharing a frame still degrade on every platform.
Yes. Hexcoded's plan comparison shows real human actors and Talking Actors starting at Starter rather than the entry plan. AI actors and your own library are available throughout.
Three tiers, one library
Platform AI actors, over 1,000 human creators with consent verified and payment on every render, or your own — generated from a description or uploaded from a photo. Everything you create saves with a handle and works in every tool.
See the actor libraryMore on characters, consent and reusable elements in Elements.