Abstract: I'll describe two cameras and a comparison of many cameras. In the random camera, we use a 'lens' which creates a pseudo-random relationship between incoming light rays and resulting sensor locations. We studied various properties (both good and bad) of the resulting camera and have built a prototype. The coded aperture camera is a conventional SLR camera but with a coded pattern of holes in the aperture. This gives a depthdependent blur which is both easy to identify and easy to deblur, allowing us to estimate, from the captured image, both an all-focus image and (roughly) the depth everywhere. Finally, we analyze cameras as linear projections of the 4-d lightfield and develop a Bayesian framework to study how well any given camera can recover the incident lightfield from its data. This gives a common framework in which to compare the performance of ordinary lenses, stereo cameras, random cameras, lenticular arrays, pinhole cameras, coded aperture cameras, etc. Joint work with Frédo Durand, Rob Fergus, Anat Levin, and Antonio Torralba.