How photo to video AI works
What the model does with your still photo, what it cannot preserve, and how to prepare a photo that gives a usable clip.
From one frame to many
A photo to video model reads a single frame, infers depth and pose, and then generates the frames that follow: the head turns a few degrees, the chest rises with a breath, the hair moves a little. Nothing in the output is filmed — every extra frame is invented, conditioned on your photo.
Because the model is inventing, the parts it has little evidence for move most freely: hands, ears, the far side of the face in a three-quarter shot, and backgrounds with repeating patterns. That is why a simple front-facing portrait usually gives a cleaner result than a dynamic action shot.
What stays and what drifts
Identity is anchored to the photo, so the face, hairstyle and clothing stay recognisable. Fine detail is not guaranteed: small jewellery, thin glasses frames, printed text on clothing and tattoos can flicker or soften over a clip.
Camera motion is the other variable. Asking for a static camera keeps the frame stable and the face consistent; asking for a push-in or orbit adds energy but gives the model more freedom to drift.
Preparing the photo
Use one person, facing the camera, with the whole head inside the frame and light coming from the front. Avoid heavy beauty filters, low-light noise, cropped chins, sunglasses, and photos where the face is turned more than about 30 degrees.
Portrait orientation crops best into a 9:16 clip. A landscape photo still works, but the model has to invent more of the frame at the top and bottom.
Start with the tool
Credits per motion are listed on the tool, packs start at $4.99.