CVPR2023

InstructPix2Pix: Learning to Follow Image Editing Instructions

Tim Brooks, Aleksander Holynski, Alexei A. Efros

摘要

Replace the fruits with cake" "Make his jacket out of leather" "Add fireworks to the sky" "Turn it into a still from a western" "Swap sunflowers with roses" "What would it look like if it were snowing?" Figure 1 . Given an image and an instruction for how to edit that image, our model performs the appropriate edit. Our model does not require full descriptions for the input or output image, and edits images in the forward pass without per-example inversion or fine-tuning.