Text2LIVE: Text-Driven Layered Image and Video Editing

We present a method for zero-shot, text-driven editing of natural images and videos. Given an image or video text prompt, our goal is to edit the appearance existing objects (e.g., texture) augment scene with visual effects smoke, fire) in semantic manner. train generator on internal dataset, extracted from single input, while leveraging external pretrained CLIP model impose losses. Rather than directly generating edited output, key idea generate layer (color+opacity) that composited over input. This allows us control generation maintain high fidelity input via novel losses applied layer. Our neither relies nor requires user-provided masks. demonstrate localized, edits high-resolution videos across variety scenes. Webpage: http://www.text2live.github.io .