We present a method of synthesising a texture using dynamic neighbourhood matching. Like existing pixel-based methods, the output is synthesised by comparing neighbourhoods of sample and target pixels. However unlike most methods, we do not specify a fixed neighbourhood size a priori. Instead the window size is determined dynamically for each pixel that is synthesised. Typically the output image is initialised with noise. In our approach, we use patches as a seed with which to guide the matching process. The patches are retrieved using the watershed method to isolate texture elements. As a result our outputs have reduced smudging, and fewer jagged drawn out texture anomalies that are problematic in existing pixel-based methods.