This paper is about scene modeling and image synthesis from existing views. From some known views of a 3D scene, called reference views, we synthesize other views of the same scene, under another point of view. The outline of the whole process is simple: we first match the reference views, then we reconstruct the matched points in projective or Euclidean space. We build the new view by projecting these points onto the plane of a virtual camera. We don't need to calibrate the cameras, as we don't need a real 3D model of the scene to synthesize the new view (a projective model is enough; transfer can be achieved by using the trilinear tensor or plain projective re-projection). This method is equivalent to a fast and realistic image synthesis. Fast, because we don't need to build a complex model of the scene by hand. Fast and realistic, because we don't have to render the synthesized view: the intensities or colors of the pixels are directly extracted from the (real) reference views. This technique has numerous potential uses in video compression, image synthesis, and more generally in all Virtual Reality applications.