Magic Hour face swap

Keep the performance the user likes and change the intended face. Accept “put me in this clip”; choose the technical settings and inspect the result within the existing budget. Use supplied authorized media, preserve delegated choices and deliver the actual file.

Choose the right change

Face Swap changes facial identity. It does not promise a new body, hairstyle, costume, scene or voice. Use character_replace_create_video for a whole-character change; use ai_image_editor_create_image to establish a new still before Image-to-Video when no source performance exists. Never create a replacement scene from text when the user asked to preserve their clip.

Inspect the source interval, face size, turns, occlusions, motion blur, cuts and other visible people. Use a clear reference of the intended identity with an unobstructed face. Match the reference angle to the difficult portion where possible; a heavily beautified reference can remove distinguishing details. Do not modify protected facial traits to chase generic attractiveness.

For a long or difficult video, render one representative interval containing a turn or obstruction before expanding within budget. A static frontal preview cannot establish reliability through cuts. Keep the original video and audio for comparison and final assembly.

Select the intended face

For one visible person, use face_swap_create_video with assets.video_source: file, assets.video_file_path, assets.image_file_path and assets.face_swap_mode: all-faces. The image is the new identity; the video contains the performance. The live description may mention source_file_path, but the actual schema's input is image_file_path.

For a specific person in a cast, use face_detection_detect_faces on the exact source asset, then retrieve that detection ID until complete. Inspect the returned face previews to identify the requested person. Use individual-faces with face_mappings: [{original_face: <returned path>, new_face: <reference path>}]. Use the returned path exactly; never invent an index or infer that “left person” is the same person across shots. Do not use all-faces when bystanders must remain unchanged. Ask only when the intended person is genuinely ambiguous.

If cropping or trimming creates a new source asset, run detection against that asset before mapping. start_seconds and end_seconds refer to the submitted video's timeline; a locally trimmed file starts at zero. Preserve output duration and sound intentionally rather than assuming trim offsets behave the same across tools.