Visual ensembles, like leaves on a tree, often share properties such as shape or color. Thisredundancy can be quantified as a feature probability distribution whose summary statistics (e.g.,mean) observers can report explicitly. Here, we show that such explicit reports underestimate therichness of ensemble perception. Participants (N=12 per condition) searched for an odd-one-outtarget among heterogeneous distractors and their memory of distractor characteristics was testedexplicitly or implicitly. Our observers could explicitly distinguish distractor sets with differentmean and variance, but not differently-shaped probability distributions. In contrast, implicitassessment revealed encoding of mean, variance, and importantly, also distribution shape.Furthermore, explicit measures had common noise sources that distinguished them from implicitmeasures. We conclude that feature distributions are encoded in rich detail and can guidebehaviour, while explicit summary judgments are based on information equivalent to a fewsamples from fully encoded distributions.