Turning a photo into video: image-to-video, explained

Text-to-video invents a clip from a description. Image-to-video starts from a picture you supply and animates it. The second gives you far more control, because the hardest part — what everything looks like — is already settled before the model runs.

Your first frame is most of the result

The model treats your image as frame one and extrapolates. Whatever is wrong in the still gets amplified in motion. In practice that means:

• Sharp beats soft. Motion blur in the source turns into smearing across the clip.
• Room to move. A subject cropped tight at the edges has nowhere to go, and the model will either squash it or drift the camera awkwardly.
• One clear subject. Busy scenes with several competing focal points tend to wobble as the model tries to animate all of them.

Prompt the motion, not the picture

The image already describes the scene. Your prompt should describe what changes: "slow push in, hair moves in the wind, clouds drift left". Re-describing the contents of the photo wastes the prompt and sometimes fights the image.

Camera language works well here too — push in, pull back, pan left, orbit, static locked-off shot. Models are noticeably better at camera motion than at complex subject motion.

What these models are still bad at

Hands doing precise things. Text staying stable across frames. More than one person interacting. Anything requiring an object to be conserved for several seconds — objects tend to melt or multiply. Keep the ask modest and the clip short and the hit rate goes up sharply.

Duration and resolution cost real money

Video is priced by length and resolution, and both scale steeply. Find your shot at the shortest duration and lowest resolution the model offers, then re-run the winner at full settings. The composer shows the credit cost updating as you change those controls, so there are no surprises.