blended-latent-diffusion

Why we included this project

Blended Latent Diffusion is the official implementation of a SIGGRAPH 2023 paper that tackles a specific editing problem: changing one region of an image while leaving the rest untouched. You give it an image, a mask over the area you want to change, and a text prompt, and it regenerates only that masked region to match the prompt. The repo includes the original Latent Diffusion pipeline plus a Diffusers-based path that works with Stable Diffusion v2.1 and SDXL, so you can choose the backbone that fits your hardware. It also has a reconstruction step that restores the background after editing, which helps avoid the artifacts you often get with masked edits. This is research code, not a polished tool, so expect to adapt it, but it is a good reference point if you are building text-guided local editing or studying how latent-space diffusion handles masked regions.