Video-to-video is a generation workflow in which an existing video clip serves as the primary input, with an AI model processing the footage to transform, restyle, or reinterpret it according to additional text or image guidance. Unlike text-to-video which generates from scratch, video-to-video uses the motion, structure, and temporal information of the input footage as a foundation, applying changes to the visual surface or style while preserving the underlying movement and composition.

Video-to-video workflows enable a range of applications including style transfer across video (applying an artistic style to live-action footage), video enhancement and restoration, changing the visual appearance of environments or subjects while maintaining their motion, and using rough or reference footage as a structural guide for generating polished AI content. The technique is particularly useful when the motion in a scene is complex and difficult to describe in a text prompt - using real or rough footage as input provides the model with precise temporal information that text alone cannot communicate. The degree to which the output adheres to the input structure versus reinterpreting it is typically controlled by a conditioning strength parameter.

Video-to-video workflows expand the creative possibilities available when working with existing footage, whether that footage is live-action material that needs visual transformation, rough proxy content shot as a motion reference, or previous AI generations that need restyling or refinement. Combining video-to-video with text prompting gives creators control over both the motion structure and the visual treatment of the final output.