In this paper we present a new pyramid architecture, which contains two approaches to reduce the turnaround time dramatically compared with conventional pyramid machines, for bottom-up image analysis. Normalized input approach alleviates the affection of objects' locations and sizes. Self-timed alternate-mode approach utilizes the advantage of picture compactness. As a result, an image-dependent pyramid model is derived. The operation of the new model is illustrated by examples of standard algorithms for interior-based computations (e.g., area) and border-based computations of local properties (e.g., perimeter).