Use a supported vision feature to analyze photos. Uploaded video needs a verified video-analysis provider; file storage alone does not add AI analysis.
Reading a photo
Where vision is enabled, the built-in vision endpoint accepts image input and a prompt. Have the builder check availability and input limits for the deployed app, then test the actual kind of image your customers will submit. Ordinary text chat is not automatically a photo-upload interface.
An AI description is not a verified detector for exact face or pet-head coordinates, size and rotation. If you need reliable automatic placement, choose and test a detector that defines those outputs for your subject.
For a simpler overlay, let the user drag, resize and rotate a sticker in the app. That can preserve the original photo and use ordinary browser code without an AI call. Generating or editing an image is a separate operation and may change details of the original.
Analyzing uploaded video
Uploading or recording a video does not make it an input to the built-in text-chat endpoint. That endpoint expects text; the vision endpoint's image support is not a video-input contract.
For a workflow review, a useful starting point is a few screenshots of key moments plus a written explanation. Test that the app's image-processing feature actually receives those screenshots. If the task requires watching the whole video, verify a specific video-analysis provider, its accepted input, limits and cost before building the upload flow around it.
Keep access to uploaded customer media restricted to the people who need it. Built-in AI processing uses the owning workspace's credits; a separately connected provider has its own billing. See Who pays for app AI and Recording video.