You give it a video and a camera path, and it returns a new clip of the same event seen from that path. You can circle around the subject, slide sideways, move closer or further away, or rise and drop. Time is a separate control: let the action play, slow it to half speed, or hold one frame still while the camera keeps moving, the frozen "bullet time" look from action films.
It also works from a single photo. Give it one image and it builds a short camera move through the scene, as if you had filmed it.
Before anything is generated, it estimates the 3D shape of the scene from your footage and shows you a rough preview from the new angle, with grey patches wherever the original camera saw nothing. The video model then fills those patches and cleans up the picture. Each take is about 3 to 10 seconds long. Meridian is not a standalone model: it is a pair of small add-on weight files (LoRA adapters) for MiniMax's H3 video model.