Leaf area is the most relevant scalar variable for describing the dynamics of vegetation on seasonal time scales, and hence is required as part of the land-surface component of a meteorological model. A mathematical scheme for the dynamic vegetation component of such a model is formulated and reduced to a toy model for seasonal leaf dynamics. Leaf growth is seen as a temperature-initiated instability of the ecosystem that drives it away from its state of winter dormancy; the onset of dormancy in autumn consists of cold temperatures breaking the summer-time attractor; the temperature-dependent controls are represented by a “ramp-up” function. Results from ensemble simulations driven by a stochastic temperature model show that leaf variability statistics have a very strong seasonality, such that variability is largely confined to spring and autumn. These variability windows promote non-Gaussian and nonstationary statistics that occur when the stable attractor of one season has flipped to the stable attractor of the other season. During such periods of high variability, any dynamic vegetation model will be most unreliable without observational constraints because of its unstable trajectory.