replace-in-video
Installation
SKILL.md
Replace in video
Swap a single on-camera element in a source video (a character, a product, a garment) for one from a reference image, while the motion, timing, camera movement, lighting, audio, and background all carry through unchanged. The lever is a source video plus one to three reference images, and a prompt that names exactly what to replace and what to preserve. This is not re-shooting or re-prompting the scene, it is a targeted swap that leaves everything else frame-faithful.
Inputs to collect
- The source video. The clip whose scene, motion, and audio you want to keep. Bring footage you have rights to (see Technique). (Ask only if none provided.)
- The reference image(s). What the new element should look like: a clean portrait for a character recast, a bare product/garment photo for an object swap. One is enough, up to three.
- What to replace and what to keep. The specific on-camera element being swapped, and whether anything ambiguous nearby (a second character, a held object) must stay.
- Optional: target resolution (720p draft vs 1080p finish), and whether the source audio should play, drive lip sync, or be stripped.
Models
- Default: Pruna P-Video-Replace (
prunaai:p-video@replace). The model for swapping one on-camera element in a clip while preserving the rest. Takes the source underinputs.videoand up to 3 images underinputs.referenceImages. Capabilityio:video-to-video+op:edit. - Reversed direction: if you want the image to be the scene and the video to only lend its motion, that is P-Video-Animate (
prunaai:p-video@animate), not this. Replace keeps the video's world, Animate keeps the image's. - To generate the source first (an iconic shot, a UGC pitch) before recasting it, route the generation to a flagship image-to-video model via
product-ad-video, then run replace on the result. - Confirm the model is
liveand inspect its exact input fields viarunware-models+runware-runbefore calling. Never hardcode a stale choice.