Survey on low-level controllable image synthesis with deep learning

Deep learning, particularly generative models, has inspired controllable image synthesis methods and applications. These approaches aim to generate specific visual content using latent prompts. To explore low-level controllable image synthesis for precise rendering and editing tasks, we present a survey of recent works in this field using deep learning. We begin by discussing data sets and evaluation indicators for low-level controllable image synthesis. Then, we review the state-of-the-art research on geometrically controllable image synthesis, focusing on viewpoint/pose and structure/shape controllability. Additionally, we cover photometrically controllable image synthesis methods for 3D re-lighting studies. While our focus is on algorithms, we also provide a brief overview of related applications, products and resources for practitioners.