We present a new programming framework, Quasar, which facilitates GPU programming. Our high-level programming language relieves the developer of all implementation details such that he can focus on the algorithm and the required accuracy. The proposed programming framework consists of a simple high-level programming language, an advanced compiler system, a runtime system and IDE. The Quasar language is a high level scripting language with an easy to learn syntax similar to python and MATLAB. This makes Quasar well suited for fast development and prototyping. A Quasar program is first processed by a front-end compiler that automatically detects serial and parallel loops that could be accelerated by heterogeneous hardware. In a second compilation phase, a number of back-end compilers processes the output of the front-end compiler, thus generating C++, OpenCL or CUDA code. Based on the generated code the runtime system can dynamically switch between CPU and GPU. This automatic scheduling at runtime is done by analyzing the load of all devices, the memory transfer cost and the complexity of the task. This way, the programmer is relieved from complicated implementation issues that are common for programming heterogeneous hardware.We validated the use of Quasar on a number of complex image processing and computer vision algorithms. These programs range from denoising to automated image segmentation applications. Using Quasar we get speed-up factors of 4 to over 60, depending on the application. All results were achieved using an NVIDIA GeForce M750.