AI spokesperson: build the avatar from a prompt, profile or photo?

For a brand spokesperson who is not a real person, start with a prompt; use profile traits when your app has to produce avatars from data; use a photo only when the face belongs to someone who has agreed. All three cost the same flat $0.95 per avatar on Sume and give you the same thing: a reusable handle that every later video names.

The difference is control and rights, not price. The details below are from the Create new avatar docs and the OpenAPI reference, read 2026-10-03.

What does each input take?

Every request carries an avatar_handle, and an Idempotency-Key header makes a retry safe when you reuse it for the same payload. A leading @ is accepted and stored without it.

Which input fits a brand spokesperson?

• A character you invent: use a prompt. You can write the look you want, then judge the result and create another handle if it is off.
• A cast you generate from a roster: use profile traits. The fixed fields make results easy to script, but the choice list is narrow: eight ethnicity values, two sex values.
• A named founder or employee: use a photo, with a written release first. See the release checklist.
• No wish to create anything: search the catalog of Sume's stock avatars, which use the reserved sume_ handle prefix, and render with one of those.

How do I keep the spokesperson consistent?

Consistency comes from reusing one handle. Every talking video names the handle, so the same face appears across a series, and each video holds one avatar. If you want different looks of the same person, Sume treats each look as its own avatar with its own handle; there is no look-pack feature, so plan the handles in advance, for example host_kitchen and host_office.

Before you commit to a spokesperson for a campaign, render a short check with the cheapest tier: a 4-second video at standard is under a dollar at $0.184 per second. A preview of first-frame stills is cheaper still to review, and it shows the face in the scene you plan to use.

What does a first test look like?

Create two avatars from a prompt, with different wording, and one from profile traits. Each is $0.95, so the whole test is $2.85. Render the same 5-second line with each at standard quality and compare. You are checking whether the face reads as a person your customers would believe in, whether it suits your product, and whether it holds up at phone size.

If none work, change the prompt rather than the tier. A higher quality tier changes the render, not who the person is. The avatar's identity is set when you create it, so a weak face is fixed at creation, not at render.

What can go wrong?

• Naming: a handle names one avatar, so give each variant its own clear handle.
• Prefix: handles starting with sume_ are reserved for Sume's own avatars and are rejected for yours.
• Photo URL: the image_url must be public HTTPS, not a signed or private link.
• No delete route: the API lists create, list and read only, so create fewer, better-named avatars.

What about rights and disclosure?

An invented face has no signer, which removes the release step, but it does not remove the label. Viewers still need to know the spokesperson is AI where the platform or the law says so, and you still cannot present a generated person as a real customer. If the invented face happens to resemble a real person, you can still have a likeness problem, so look at the faces you generate before they run. Avatar 1.0 speaks English only in code, so a spokesperson for other markets needs a different route.