The spread and coloration of pair-wise panned virtual sources are investigated in listening tests. The results are compared with results from corresponding simulations using a binaural auditory model. The results show that amplitude panned virtual sources produce highly conflicting cues at frequencies between 700 Hz and 2 kHz, and that they are localized according to different cues depending on signal attributes. A DSP-based method for enhancing the virtual source quality is proposed.