The principle of efficient coding suggests that processing in the early visual system should be optimized and adapted to the environmental statistics. An intracellular study of the primary visual cortex (V1) in the anesthetized and paralyzed cat showed that the reliability of the neural response is optimized for natural statistics. Using the same natural and artificial stimuli, we recorded the neuronal population activity (single unit, multi-unit and local field potentials) in cat’s V1 with high-density linear silicon probes. We first investigated the reliability and of the mesoscopic signal with the intracellular signal and explored its laminar dependency. Our results showed that natural images evoke, at all scales, the most reliable response, suggesting that V1 is better suited to efficiently encode natural statistics. In addition, granular and infragranular layers displayed higher reliability levels than the supragranular one. This argues for a functional filtering of the pertinent information between these layers. We also explored which statistics of the natural images produce this reliable response. Finally, we specifically addressed the role of the correlations between neurons (within and between layers) by measuring the amount of shared variability and signal of the neuronal population in response to our stimulus set. We observed that natural images always evoked higher correlations. We did not observe a strong decorrelation at the single cell level but instead at the scale of groups of neurons, with those that are close together being more correlated and farther apart less correlated, arguing for a functional clustering of the neurons into coherent “neural mass”.