The Ask AI pipeline
Staff only
Internal design notes — stripped from the public build. The public pages describe Ask AI only as observable behaviour ("suggests a designed avatar from your photo"); the mechanism below stays here.
Ask AI turns a photo into a designed avatar:
- An on-infrastructure vision model reads coarse descriptors from the photo — skin tone, hair, glasses, facial hair, and so on. The photo is discarded, never stored.
- Portrait's own deterministic map turns those descriptors into
avataaarsoptions. - Those options seed the Design editor. The user tweaks and saves through the normal designed path.
It's assisted, not an exact likeness. The value — and the thing worth not publishing — is the
descriptor set and the deterministic descriptor → avataaars option map, which together are what make
the suggestion reproducible without a per-request model call at read time.
Endpoint: POST /me/avatar/from-photo.