With the ever increasing costs of manual content creation for 3D virtual worlds, the potential of generating content automatically becomes too attractive to ignore. However, for most designers, procedural generation methods are complex and unintuitive to use, and offer little user control. Furthermore, due to their specialized nature, separately generated results are not easily integrated into a complete and consistent virtual world. In this thesis, we propose declarative modelling of virtual worlds, an approach that enables designers to concentrate on what they want to create instead of on how they should model it. To realize this approach, we have devised a framework, building upon proven results on procedural generation, constraint solving and semantic modelling. The foundation of this framework is provided by a semantic model for virtual worlds, which structures terrain features in several levels of abstraction, from a coarse user specification to concrete 3D geometry, and enriches objects with relevant information on their functionalities, services and roles. The framework supports structured integration of procedural methods at different levels of abstraction. With these integration methods embedded in our framework, we are able to harmonically apply existing procedural methods in combination to generate complete virtual worlds with detailed objects. We allow for intuitive interaction with the framework, providing user control at several levels of granularity. Our interaction methods, such as procedural sketching, can be freely mixed and are interactively evaluated, enabling a short feedback loop. Each of these methods has its own added value and they complement each other. In order to form a consistent and plausible environment, generated features also have to be properly embedded in the virtual world. To this end, we introduced automatic consistency maintenance, which uses generic methods to handle any interactions that occur between features. This removes a huge burden from the designer, who is now freed from the task of continuously fitting all content together and keeping the world consistent. We believe that the combination of these contributions, as successfully implemented in our prototype SketchaWorld, significantly helps designers in the process of generating virtual worlds.