The extent to which so-called low-level features are relevant to predict gaze allocation has been widely studied recently. However, the conclusions are contradictory. Edges and luminance contrasts seem to be always involved, but literature is conflicting about contribution of the different spatial scales. It appears that experiments using an-made scenes lead to the conclusion that fixation location can be efficiently discriminated using high-frequency information, whereas mid- or low frequencies are more discriminative for natural scenes. This paper focuses on the importance of spatial scale to predict visual attention. We propose a fast attentional model and study which frequency band predicts the best fixation locations during free-viewing task. An eye-tracking experiment has been conducted using different scene categories defined by their Fourier spectrums (Coast, Open-Country, Mountain, and Street). We found that medium frequencies (0.7-1.3 cycles per degree) globally allowed the best prediction of attention, with variability among categories. Fixation locations were found to be more predictable using medium to high frequencies in man-made street scenes and low to medium frequencies in natural landscape scenes.