Magnetic resonance imaging (MRI) offers an excellent opportunity to study human body composition. Differentiation of fat and nonfat tissue allows quantitative analysis of adipose tissue volume from MRI images. Cylindrical phantoms containing various proportions of water and oil were scanned using a spin echo sequence with an echo time of 12 ms, a repetition time of 150 ms with 20 mm slices. A previously described method of defining a fat pixel threshold underestimated the known volume of fat in the phantoms by a mean of 9.7%. Eight young male volunteers (mean age, 22 years) were scanned twice in succession using the same protocol as with the phantoms. The volume of fat calculated from images of eight subjects was upwardly adjusted by the amount the pixel thresholding method underestimated the known volumes of fat in the phantoms. The fat volume determined from the images of human subjects was converted to weight and then percent body fat (BF). The mean percent BF of the two MRI scans of the eight subjects was 17.66% (SD±5.43), which was significantly greater (p = 0.003) than 15.87% BF (SD±4.73) determined by hydrodensitometry. When fat volume was not corrected for the pixel thresholding method's underestimation of the known volumes of fat in phantoms, the MRI percent BF (15.96%) was not significantly different from hydrodensitometry. Test‐retest MRI total body fat volume correlated highly (r = 0.99). Imaging of the total body was brief and subjects incurred no discomfort. Before MRI becomes a standard in body composition research, the problems associated with image acquisition (movement artifact and ghosting) and image analysis (validation of methods, accounting for artifact, ghosting, and nonfat pixels appearing as fat pixels) require further attention. We suggest a combination of pixel thresholding and visual verification to segment fat and nonfat tissue.