Counting colony-forming units (CFUs) in Petri dishes is an essential task in the food, cosmetic and pharmaceutical industries. Because automating it with a detection model requires hundreds, if not thousands, of labelled images that are often unavailable in analysis-laboratory workflows, a training-free pipeline is proposed that turns a single labelled image into a fully labelled synthetic dataset, balancing realism and diversity.

Starting from a single labelled Petri dish plate, the pipeline uses pretrained diffusion models with IP-Adapter and ControlNet to remove its colonies and obtain an empty dish, generate varied empty dishes and colonies, and compose new labelled plates. The generated images are evaluated both for their realism against real plates and for their ability to train a YOLOv8-s detector, measured on the public AGAR dataset.

A detector trained on generated images achieves a mAP@0.5 of 0.882, approaching 0.985 with 350 real annotated plates and outperforming the Pawlowski-inspired baseline (0.775) and the single plate alone (0.616).

From a single labelled image and with no training nor additional fine-tuning, the generated synthetic data trains CFU detectors approaching full-supervision performance.