TF-ICON

Why we included this project

TF-ICON is the official ICCV 2023 implementation of a training-free method for cross-domain image composition. The idea is simple: you give it a background, a foreground object, its mask, and a placement mask, and it inverts the real images into latent space and blends them through a pretrained Stable Diffusion checkpoint, with no finetuning or per-instance optimization. That training-free design is what makes it worth a look for anyone doing image editing with diffusion models. The repo also introduces the 'exceptional prompt' trick for more accurate real-image inversion, which is independently useful in other editing workflows. It is a research codebase rather than a turnkey product, so expect to run scripts and manage model weights yourself, but the setup is documented and benchmark data is included for reproducing the paper's results.