Millions of individuals explore virtual worlds every day, for entertainment, training, or to plan business trips and vacations. Video games such as Eve Online, World of Warcraft, and many others popularized their existence. Sand boxes such as Minecraft and Second Life illustrated how they can serve as a media, letting people create, share and even sell their virtual productions. Navigation and exploration software such as Google Earth and Virtual Earth let us explore a virtual version of the real world, and let us enrich it with information shared between the millions of users using these services every day. Virtual environments are massive, dynamic 3D scenes, that are explored and manipulated interactively by thousands of users simultaneously. Many challenges have to be solved to achieve these goals. Among those lies the key question of content management. How can we create enough detailed graphical content so as to represent an immersive, convincing and coherent world? Even if we can produce this data, how can we then store the terra–bytes it represents, and transfer it for display to each individual users? Rich virtual environments require a massive amount of varied graphical content, so as to represent an immersive, convincing and coherent world. Creating this content is extremely time consuming for computer artists and requires a specific set of technical skills. Capturing the data from the real world can simplify this task but then requires a large quantity of storage, expensive hardware and long capture campaigns. While this is acceptable for important landmarks (e.g. the statue of Liberty in New York, the Eiffel tower in Paris) this is wasteful on generic or anonymous landscapes. In addition, in many cases capture is not an option, either because an imaginary scenery is required or because the scene to be represented no longer exists. Therefore, researchers have proposed methods to generate new content programmatically, using captured data as an example. Typically, building blocks are extracted from the example content and re–assembled to form new assets. Such approaches have been at the center of my research for the past ten years. However, algorithms for generating data programmatically only partially address the content management challenge: the algorithm generates content as a (slow) pre–process and its output has to be stored for later use. On the contrary, I have focused on proposing models and algorithms which can produce graphical content while minimizing storage. The content is either generated when it is needed for the current viewpoint, or is produced under a very compact form that can be later used for rendering. Thanks to such approaches developers gain time during content creation, but this also simplifies the distribution of the content by reducing the required data bandwidth. In addition to the core problem of content synthesis, my approaches required the development of new data-structures able to store sparse data generated during display, while enabling an efficient access. These data-structures are specialized for the massive parallelism of graphics processors. I contributed early in this domain and kept a constant focus on this area. The originality of my approach has thus been to consider simultaneously the problems of generating, storing and displaying the graphical content. As we shall see, each of these area involve different theoretical and technical backgrounds, that nicely complement each other in providing elegant solutions to content generation, management and display.