A claimed limitation of current AI video/image models: continuity of spatial position, lighting, and geography across separate generations is not tracked by the model at all — the creator must enforce it manually, shot to shot. Presented as the root problem motivating two separate mitigation techniques: Video-Sequence Generation for Frame-to-Frame Consistency (generating a continuous video rather than separate stills to preserve lighting/detail within a shot) and Blocking Lock via Position Reference (external position references to stop objects/characters from teleporting between separate shots).
Из тем: Holding Continuity Across Independently Generated Shots