Image Editing with Deep Generative Networks – IDeGeN

The goal of this project is to use generative neural networks in order to perform image editing, and in particular the editing of "high level" attributes, such as the expression or appearance of a face. This is especially necessary for the film post-production industry. Currently, generative networks achieve impressive image synthesis results, thanks to random draws in "latent" spaces, learned on data. We will therefore seek to edit in these spaces. However, these do not currently allow the fine control of attributes necessary for editing. This project therefore seeks to impose useful properties on latent spaces, such as a good separation and organization of attributes. We will propose algorithms for projection and navigation in the latent spaces of generative models. In particular, we will be interested in the disentanglement and linearity properties of these spaces. We will seek either to establish our own architectures and cost functions to create latent spaces with good properties, or to use a posteriori powerful existing models, and / or to analyze the heuristics that these use. We are also interested in the problem of localized editing (in an image), which is important for digital artists. The main application is the editing of photographic images, but we also consider the case of medical images, where we will seek to understand these medical images by the bias of generative models and by the latent spaces of these. The project will have impacts on the fields of film post-production, which is very expensive and represents an extremely laborious task, and medical imaging, where it is relevant to offer analysis tools that allow comparison of patient conditions or potentially predict certain pathologies.