blended-diffusion

Why we included this project

Blended Diffusion is the official implementation of a CVPR 2022 paper that introduced a way to edit just one region of a photo using a text prompt. You give it an image, a mask over the area you want to change, and a short description; a CLIP model guides a diffusion process so only that masked region is regenerated, leaving the rest of the photo intact. That makes it a useful reference for anyone building or studying image-editing systems where background fidelity matters. The repo includes a working command-line interface, pretrained checkpoints, and example scripts for common edits like adding an object, replacing one, swapping the background, or extrapolating the image. The README also points to the follow-up Blended Latent Diffusion, which runs faster and produces better results, so this is a good starting point before moving on to that.