cycle-diffusion
CycleDiffusion is an official PyTorch implementation of the ICCV 2023 paper A Latent Space of Stochastic Diffusion Models for Zero-Shot Image Editing and Guidance by Chen Henry Wu and Fernando De la Torre from Carnegie Mellon University. This software formalizes the concept of random seeds in diffusion models, providing a method to infer these latents from real images. It enables zero-shot image-to-image translation using text-to-image models like Stable Diffusion without requiring paired training data. The tool allows users to translate an input image x from a source text t to a target text t-hat while preserving the underlying visual content consistent with the original image. Key features include unpaired image translation between related domains and seamless integration with Hugging Face Diffusers for streamlined pipelines. The repository includes code for customized usage, evaluation on standard data, and access to pre-trained diffusion models. It supports applications such as style transfer, object repl