We conducted a series of flume experiments to investigate the response of self-formed gravel-bed channels to floods of varying magnitude and duration. Floods were generated by increasing the discharge into a channel created in sand- and gravel-sized sediment with a median grain size of 2 mm. Flooding increased the Shields stress along the channel perimeter, causing bank erosion and rapid channel widening. The sediment introduced to the channel by bank erosion was not necessarily deposited on the channel bed, but was rather transported downstream, a process likely facilitated by transient fining of the bed surface. At the end of each experiment, bank sediments were no longer in motion, “partial bed load transport” characterized the flat-bed portion of the channel, and the Shields stress approached a constant value of 0.056, about 1.2 times the critical Shields stress for incipient motion. Furthermore, the discharge was entirely accommodated by flow within the channel: the creation of a stable channel entirely eliminated overbank flows. We speculate that similar processes may occur in nature, but only where bank sediments are non-cohesive and where channel-narrowing processes cannot counteract bank erosion during overbank flows. We also demonstrate that a simple model of lateral bed load transport can reproduce observed channel widening rates, suggesting that simple methods may be appropriate for predicting width increases in channels with non-cohesive, unvegetated banks, even during overbank flows. Last, we present a model for predicting the equilibrium width and depth of a stable gravel-bed channel with a known channel-forming Shields stress.