You give it one photo, and it returns a short clip in which the camera travels a full circle around the subject while time stands still. A dancer stays mid pose, hair stays in the air, and only the viewpoint moves. The last frame is your original photo again, so the clip loops cleanly or joins onto another one without a visible jump.
That return trip is the hard part. Asked to start and end on the same image, the H3 video model on its own barely moves the camera; given only a starting image, it drifts off and never comes back. This add-on, a small set of extra weights known as a LoRA, teaches it to make the whole lap. If you read about Meridian in issue 115, which re-shoots a scene from a new angle, this one does a single job: the complete 360 degree turn.
It was trained on just 28 orbit videos rendered around 3D captures of people, so it is at its best with one person in the frame. Each clip runs about three seconds.