The utility of PET and PET/CT in research and diagnosis of cancer, cardiac and neurological disorders has been widely demonstrated. Phantoms with well defined geometries that accurately model radiolabeled tracer concentrations and photon attenuation coefficients are suited for characterization of imaging systems, but not as well suited for evaluating methods sensitive to detailed anatomical structure, such as algorithms for monitoring tumor response. An ideal phantom would have the shape and activity distribution of a realistic tumor and would be useful in evaluation of automated image analysis systems. Such a phantom, imaged at sites involved in clinical trials, would be valuable for evaluating consistency and accuracy. We have developed a method of creating such phantoms by incorporating radioactive tracer as dye for a cellulose powder based rapid prototyping system. This allows us to create phantoms with spatial resolution limited only by the stereolithography printer system (slice thickness is 0.18 mm, printing resolution is 600 dpi). We have evaluated the method by printing several small phantoms with 18F and measuring activity in a gamma counter. The relative standard deviation of the activity of multiple identical phantoms was 2%. Activity in unlabeled parts was less than 2% of adjacent labeled regions. We have created and printed realistic phantoms based on the SPL human brain atlas [1,2], the Paxinos & Watson rat brain atlas  and from PET/CT images of human lung nodules, showing that this is a practical method for making complex radioactive phantoms that model real anatomy. We are proceeding with further development to allow us to produce phantoms with multiple activity concentrations, tunable photon attenuation coefficients and long lived isotopes.