The recipe
- The model encodes your clip into a compressed representation capturing motion and layout
- Your text prompt is encoded as a description of the target style
- Starting from noisy versions of the video, the model repeatedly denoises while being nudged by both the layout and the prompt
- The result is decoded back into frames
Why layout is preserved
The encoded structure of your clip acts as a guide, so edges, positions and movements remain close to the original.
Why detail is lost
The compressed representation cannot hold every pixel, so tiny text and fine texture are the first to go.
Practical implications
- Large clear shapes survive; small details do not
- Prompts steer style but not exact geometry
- Longer clips are harder to keep consistent