Portrait Reification with Generative Diffusion Models

An application of Generative Diffusion Techniques for the reification human portraits in artistic paintings is presented. By we intend transformation painter’s figurative abstraction into a real face. The exploits recent embedding technique Denoising Implicit Models (DDIM), inverting generative process and mapping visible image its latent representation. In this way, can first embed portrait space, then use reverse diffusion model, trained to generate faces, produce most likely approximation portrait. actual deployment involves several additional techniques, mostly aimed automatically identify, align, crop relevant portion face, postprocess generated order enhance quality allow smooth reinsertion original painting.