Design by demonstrations is a promising approach for automatically generating control software for robot swarms. In this approach, users demonstrate a desired spatial organization of the swarm, and an automatic method generates control software that enables the robots to imitate it. Commonly, demonstrations are mapped into arrays of features selected through expert knowledge (e.g., relative positions between robots and landmarks). We show that simple image demonstrations are sufficient to describe the desired organization, avoiding the need for expert feature mapping. We conduct a design process that takes images of a desired swarm organization as input and applies inverse reinforcement learning with automatic modular design to produce the control software. We evaluate this method in simulations with swarms of e-puck robots. Results show that the image-based process achieves performance comparable to methods based on predefined expert features.