Image embedding for denoising generative models

Abstract Denoising Diffusion models are gaining increasing popularity in the field of generative modeling for several reasons, including simple and stable training, excellent quality, solid probabilistic foundation. In this article, we address problem embedding an image into latent space Models, that is finding a suitable “noisy” whose denoising results original image. We particularly focus on Implicit Models due to deterministic nature their reverse diffusion process. As side result our investigation, gain deeper insight structure models, opening interesting perspectives its exploration, definition semantic trajectories, manipulation/conditioning encodings editing purposes. A property highlighted by research, which also characteristic class independence representation from networks implementing other words, common seed passed different (each trained same dataset), eventually identical images.