DragDiffusion

Why we included this project

Photo editing often comes down to masking and repainting, but DragDiffusion offers a more direct approach: place a handle on the image, pull it to a new spot, and the diffusion model redraws the content so the change fits naturally. On real photographs it produces far more coherent results than the older DragGAN method. The repository contains the official code from the CVPR 2024 paper, plus DragBench, a benchmark with code and data for evaluating point-based editing techniques, which is useful when you are comparing approaches or writing your own tests. It assumes some comfort with diffusers, LoRA fine-tuning for edits on real images, and a capable GPU, so it suits researchers and engineers who want to reproduce the work or build an interactive editing pipeline, not casual users. Because this is research code with a disclaimer rather than a polished product, plan to adapt it yourself before relying on it in production.