Building a photo-to-video window scene without a prompt box

I have been building a small photo-to-video product called Burger Window AI. The whole product is one scene: upload a bright, front-facing photo of a person or a pet, and it returns a vertical clip of that subject at a night window, mid-bite, staring at the camera, then the curtain closes.

General video models ask the user to write a prompt. Most of the people this is for are posting to TikTok, Reels, or Shorts. They already have the photo. They do not want to describe lighting, camera, or timing.

So the generator is a fixed scene, not an open prompt:

No prompt writing. The input is the photo.

The pipeline is picky about the source frame:

If the photo is a side profile or a dark room shot, the window scene falls apart because the model has to invent a face it never saw.

The clip is sized for short-form feeds. The beat is simple on purpose: the freeze mid-bite is the hook, the stare is the hold, the curtain is the ending. That is the whole edit. People can post it as-is.

This is not a general text-to-video studio. One scene, one upload, one MP4. The tradeoff is obvious: you cannot ask it for a different location or a different action. If you want the window scene, that constraint is the product.

If you are building something similar, a narrow scene with a strict photo spec is easier to ship than a prompt box that can do anything and usually does it wrong.