Bystander faces in a scene photo: check likeness before image_url
A photo scene can carry a stranger's face into the render. Check every face before you send image_url; the docs do not describe a screening step.

If you use a photo as the scene for an avatar video, check every face in it before you send the URL: crop or retake the photo if a bystander appears. The Sume docs describe scene: { "type": "photo", "image_url": "..." } as an input and say nothing about screening people in it, so the check is your job.
Where a stranger's face can get in
The avatar API takes a photo as a scene reference, and it can take product images and per-scene backgrounds. A photo of a store, an office, a trade-show booth, or a street often has people in the back. A render builds a video around that image, so a bystander may appear in the final frames. That is a likeness problem even if they are small in the shot.
| Check | Look for | Fix |
|---|---|---|
| Faces | Anyone besides your avatar | Crop, blur or retake |
| Screens | Names, chats, email in view | Retake with a clean screen |
| Badges and plates | Readable names or numbers | Retake |
| Reflections | Windows and mirrors | Change the angle |
| Rights | You own or may use the photo | Use your own shot |
Use the preview
Sume's preview route produces a first-frame still before the full render, and the docs say it refreshes the stills and can include the scene image URLs. A reviewer can look at the still for faces before you pay for a render. That is a better check than a quick look at the source photo, because it shows the composed frame.
import json
body = {
"avatar_handle": "store_host",
"aspect_ratio": "9:16",
"quality": "standard",
"scene": {"type": "photo", "image_url": "https://example.com/store-front-clean.jpg"},
"script": "Welcome to our new location. Come see what is on the shelves this week.",
}
print(json.dumps(body, indent=2))A short review routine
Make the check a step in your workflow, not a habit. One person drops the photo into a shared folder, a second person looks at it at full size, and only then does the URL go into a request. Zoom to 200 percent and scan the edges, windows, and reflective surfaces, because that is where small faces hide. After the preview still arrives, look again, since the composed frame can crop or scale the scene differently from the source file.
If you cannot verify who is in a photo, do not use it. A fresh photo of an empty space costs a few minutes, and a complaint from someone who finds their face in your video costs much more.
Process notes
- Keep a folder of approved scene photos with a note that no bystanders appear.
- Do not use a photo from a public feed or a customer's post without their permission.
- If an image has a bystander, retake it at a quiet hour; blur is a last resort because it looks odd.
- A product image is not a person, but a hand or a face holding it is.
Limits
The docs do not describe a face-screening step, so do not assume one exists. Likeness rules vary by country and by use, and people can object to being in advertising even when the photo was taken in public. When in doubt, retake the photo.
Sources
Related posts
More in Sume Avatar 1.0
- Does an AI avatar presenter make a Short original?
An avatar is a delivery method; YouTube's pages ask for original substance. How to give an Avatar 1.0 Short your own angle within its 60-second job limit.
- Full-duplex avatar listens while it speaks: script a clip instead
A full-duplex avatar hears you mid-sentence. If your content is scripted, build similar beats into one Sume avatar clip with silence scenes. Curl included.
- Full-duplex or script-driven avatar: six questions before you pick
Tavus Griffin-Lite is a full-duplex video conversation model in research preview. Six questions that separate it from Sume Avatar 1.0 talking video.
- Holiday ad captions on Avatar 1.0: slam, punch or tiktok-green?
Avatar 1.0 burns captions inline in slam (default), punch or tiktok-green. No separate billed caption job, and a caption failure keeps the clean video.
Written by Sume