ReNoise: Real Image Inversion Through Iterative Noising

arXiv:2403.14602v1 Announce Type: cross
Abstract: Recent advancements in text-guided diffusion models have unlocked powerful image manipulation capabilities. However, applying these methods to real images necessitates the inversion of the images into the domain of the pretrained diffusion model. Achieving faithful inversion remains a challenge, particularly for more recent models trained to generate images with a small number of denoising steps. In this work, we introduce an inversion method with a high quality-to-operation ratio, enhancing reconstruction accuracy without increasing the number …

abstract arxiv capabilities challenge cs.cv cs.gr cs.lg diffusion diffusion model diffusion models domain eess.iv generate however image images iterative manipulation text through type unlocked