Extracting Edit-Friendly Noise Maps for Diffusion Models to Enable Diverse Image Manipulations

We present an alternative latent noise space for denoising diffusion probabilistic models (DDPMs) that enables a wide range of editing operations via simple means. Our inversion method extracts noise maps that are distributed differently from those used in regular sampling, and are more edit-friendly. This allows diverse editing of real images without fine-tuning the model or modifying its attention maps.