We evaluated orientation discrimination thresholds using an external noise paradigm. Stimuli were spatiotemporal Gaussian patches of 2D orientation noise band-pass filtered in Fourier domain. Orientation acuity was measured for various combinations of stimulus spatial bandwidth, spatial frequency, and size as a function of orientation bandwidths of the stimuli. Stimulus contrast was matched in multiples of detection threshold. Consistent with the idea that stimulus orientation bandwidth acts as a source of external noise, orientation discrimination thresholds increased monotonically in all conditions with stimulus bandwidth. To interpret these results quantitatively, we first fitted a variance summation model to the data and derived the internal orientation noise, relative sampling efficiency, and orientation tuning of the mechanism underlying orientation discrimination. Due to the equivocal biological nature of these parameters for orientation discrimination, we investigated, with a modeling approach, how neural detectors characterized by a broad orientation tuning may support orientation discrimination. We demonstrated using the ideal-observer theory that while linear models, based on either unimodal filtering or center-surround opponency, predict the monotonic relationship between orientation discrimination threshold and orientation noise, a nonlinear model incorporating a broadband divisive inhibition in the orientation domain is a better candidate due to its contrast invariance. This model, using broad and similar orientation tuning for its excitatory and inhibitory inputs, accounts for the acute orientation acuity of human vision and proves to be robust despite the variance found in natural stimuli.