pruna / p-video-replace

## Overview P-Video Replace is a Pruna video model for identity replacement in an existing clip. You provide a driver video that supplies the motion, timing, and scene context. You also provide 1–3 reference images of the subject you want inserted. The model renders a new MP4 where the reference subject appears in the scene, which helps you avoid manual compositing work. ## What you can build - Swap a spokesperson into an existing promo clip while keeping the original camera motion - Create consistent creator variants for the same driver video using different reference subjects - Turn a single performance clip into multiple character versions for storyboards - Produce identity-based video personalization when you already have the final edit - Make quick “what if this person was in this scene” creative tests for campaigns ## Inputs - A source RGB video in MP4 format. This clip provides the motion and can provide the audio track. - One to three reference images of the subject to insert. Use JPG, JPEG, PNG, or WebP. - An optional instruction text that explains how the subject should be placed in the scene. - A target output resolution choice of 720p or 1080p. - An optional target frame rate choice. You can keep the original FPS or set a fixed FPS. - An optional turbo mode toggle. This trades some visual quality for faster generation. - An audio handling choice. You can keep the source audio in the final output. - An optional “ignore audio during generation” toggle. This affects conditioning, not whether audio is saved. - An optional numeric seed for repeatable results. Use 0 to randomize. - An optional safety checker toggle that can be turned off for the generation. ## Outputs The model returns a rendered MP4 video. The output keeps the driver clip’s motion and scene timing, while inserting the reference subject into the video. The exported video matches the selected resolution and frame rate settings. If you choose to keep audio, the MP4 includes an audio track from the source clip. ## Limitations - Identity can drift when the driver clip has heavy occlusion, motion blur, or extreme head turns. - Very fast action, dense crowds, or rapid cuts can reduce subject consistency. - Strong compression artifacts, low light noise, or low-resolution inputs can cause unstable textures. - Reference images that don’t match the driver clip’s angle and lighting can look pasted in. - Background details near the replaced subject can warp if edges are ambiguous. ## Safety & compliance - Only use videos and reference images you own or have rights to use. - Get consent before generating videos of real people, especially for advertising or public release. - Don’t use this model for deception, impersonation, or non-consensual deepfakes. - Keep the safety checker on when you need stricter filtering for user-generated content.