Are you by any change using "Float" or "Half Float" for the Viewer GL Buffer?
Because those two modes don't have error diffusion. Thus they show a "banding" due to 8bit displays nature that it's not present on the float data that is being processed.
Aside that, the only reason I see to have that behavior is by having something in your pipe or connected as viewer input process that is clipping data in a really bad way.