Abstract : Video games and visual effects have always the need for more realistic images of natural scenes, including clouds. Traditional computer graphics methods involve expensive computations which currently limits their realism. In this thesis we propose new approaches for the rendering of clouds that are both realistic and fast. These goals are achieved using a methodology relying on both physical and phenomenological approaches. In the first part of this thesis we study the clouds from the point of view of a human observer, which allows us to identify the important visual features that characterize clouds. These features are the ones we seek to reproduce in order to achieve realism. In the second part we conduct an experimental study of light transport in a slab of cloud using the laws of radiative transfer. This study allows us to better understand the physics of light transport in clouds at a mesoscopic and macroscopic scale, to find the origin of the visual features of clouds, and to derive new, efficient phenomenological models of light transport. In the third part we propose two new methods for the rendering of clouds based on the findings of our study. The first model is aimed at stratiform clouds and achieves realistic rendering in real-time. The second model is aimed at culumiform clouds and achieves realistic rendering in interactive time. These models are optimized for graphics hardware. The quality of our results and models are discussed and we sketch directions for future research that can improve both the speed and the realism of our methods.