Recent improvements in processor and graphics power mean researchers can now interact with their running simulation at the same time as they view the results. This so-called 'computational steering' brings greater insight to the investigation process and is even more compelling when immersed in the visual representation of the data. Visualization of pre-computed data likewise benefits if we can generate the image without the need for separate dials and sliders. This image-based interaction, both for computational steering and visualization purposes, is therefore an important goal for scientists and engineers but typically requires specialist programming of the components. This paper introduces a new architecture that enables Modular Visualization Environments (MVEs) -- general-purpose, extensible systems -- to operate in this way. Novel elements include an input-describing data structure to convey information about the simulation parameters, and proxy graphics objects that convert ordinary image geometry into interactive elements. To promote its adoption, the architecture is designed such that the usual MVE appearance is retained. Its implementation in IRIS Explorer and Open Inventor is described and three case studies are presented: two deal with computational steering, whilst the third shows how a visualization technique, the contour plot, can be modified to take advantage of image-based interaction.