Converting a still photo into a short video looks like a model choice problem. After building one, I think it is a motion design problem wearing a model costume.
The preview lies about the hard part
Every demo shows a clean subject on a clean background. Real inputs are a person at a slight angle under mixed lighting. The model handles the demo; the failure mode is always the same: the first half-second looks right and then the motion drifts.
What helped:
- Shorter clips. Three seconds hides artefacts that eight seconds exposes.
- Explicit motion intensity. Left implicit, the model either freezes or hallucinates.
- A repeatable input checklist — lighting direction, subject angle, what is in the background.
The guide is more valuable than the button
When I looked at support questions, almost none were about generation failing. They were about which photo to use. That is a documentation problem, not a model problem.
So the highest-leverage page on the site became a short guide: how to pick a source photo, what the template does, and what it does not do. It answers the question people had before they clicked generate.
Be explicit about consent and limits
Generating a video of a real person raises obvious questions. Stating plainly that the upload must be a photo of consenting adults, and what happens to the file afterwards, removed more hesitation than any feature.
If you are documenting one of these tools, write the burning bridges video guide before you write the changelog. The instructions are the product.
Written after watching users fail on input selection far more often than on generation.
Top comments (0)