The best photo to use for an AI video
The best photo for an AI video shows one subject alone, head to toe with a margin around it, upright in even indirect light, limbs away from the body, against a plain background.
Framing, which matters most
The whole subject inside the frame with room at the top and bottom. A subject cropped at the ankles forces the model to invent feet in frame one and keep inventing them for the whole clip.
One subject. A second body in the background will be pulled into the generation, and the two will trade features.
Light, which matters nearly as much
Flat, even, indirect light. Overcast outdoors is close to ideal; so is a room with a big window and no direct sun.
What to avoid is a hard shadow across the face or body. The model reads a shadow edge as a feature and carries it through every frame, so a shadow across a cheek becomes a permanent mark that moves with the subject.
Pose and background
Upright, facing the camera, limbs clear of the torso. Arms crossed and hands in pockets are the two poses that reliably produce bad hands, because the model has to invent the hidden hand before it can move it.
A plain wall beats a busy room. Backgrounds are regenerated along with the subject, and a detailed one visibly drifts over a long clip.
Photographing a pet or a toy
Get the camera down to its level rather than shooting from standing. A photo taken from above foreshortens the legs, and the model has to guess at the parts the angle hid.
For a plush toy or a figure, stand it up and give it a plain surface. A toy lying on a patterned duvet is two problems at once.
What does not matter
Resolution, past a point. A normal phone photo has far more detail than the model needs, and a 48-megapixel file will not outperform a well-lit 12-megapixel one.
Filters, mostly — though heavy skin smoothing removes the texture that makes a generated face look like a face, so turn that one off.