Generation modes

    Segmentation

    Also called Video segmentation, Masking.

    Segmentation labels which pixels in each frame belong to a chosen object, producing a mask that other tools then act on.

    A mask is not an edit. It is a stencil: a black-and-white record of where a thing is, frame by frame. Everything interesting happens afterwards, when something else uses that stencil to delete, replace, blur, recolour or track the region it covers.

    You point at what you want rather than describe it. Implementations take a click, a box drawn around the subject, or occasionally a text label, and propagate that selection forward through the clip. Propagation is the hard part — the object turns, is occluded, leaves frame and comes back.

    It underpins several things that look like separate features. Background removal is segmentation of the subject plus a delete. Object replacement is segmentation plus inpainting. Tracking a logo across a shot is segmentation with the mask handed to a compositor.

    In practice

    • Selections are propagated, not re-detected per frame, so an early bad selection poisons the whole clip.
    • Occlusion is the standard failure: a hand passing over the subject often splits the mask.
    • Masks compose — several objects can be segmented separately and combined.

    The mistake to avoid

    Expecting a mask to be an output you can publish. It is an intermediate; on its own it is a silhouette.

    Where you will run into it

    Related terms

    The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.