From forty listing photos to a narrated tour, rendered in the browser
Estate agents already have the photos. What they don't have is the hours to cut them into a video with the right camera moves, a voiceover and captions in two formats.
The challenge
A property tour video used to mean an editor, a timeline and an afternoon per listing. The decisions an editor makes are repetitive and learnable: which room is this, which camera move suits it, what does the narration say. The expensive part was never judgement. It was assembly.
The system
- Photos
- Vision model
- Jev picks the room
- One shot per room
- Voice and captions
- Rendered in the browser
What we implemented
- Room reader
- A vision model describes each photo in words, so the next step reasons about text rather than pixels.
- Jev, the room decision
- A small decision model picks the room type from a fixed list, with a probability for each option.
- Shot director
- Gives every room the move that suits it: a lateral move along a kitchen counter, a push-in on a bedroom, nearly still on a close-up.
- Narration
- Writes and voices the script, then times the captions to it.
- Browser renderer
- Composites clips, voice, captions and motion graphics on a canvas in the customer's browser and saves the finished file.
The result
A finished tour in social and portal formats from photos the agent already had. The free tier needs no account and no key, and the finished video never sits on our servers.