Compression algorithms using floating point have always been flaky though. Need a fixed-point image model.