Modelling of physical processes such as ablation or runoff at continental or global scales provides a key challenge: a high degree of abstraction is required in order to minimize computational demands, while spatial and temporal variability of key processes, often at the sub-scale level, need to be adequately captured and reproduced within a lower resolution model. For some approaches, such as temperature index models, downscaling to lower resolutions is straightforward. However a key issue when using these downscaled models is to assess the impact of scaling on model behaviour and results, including the associated uncertainties. We assess the impact of scaling on both a simple and an enhanced temperature index melt model from 100 m to 1, 5 and 10 km resolutions. Different sub-grid parameterization approaches are applied to both models across all resolutions and tested for their suitability against high-resolution reference data, with the aim of developing a robust, scalable and computationally undemanding parameterization. Results show patterns of over- and underestimation of potential melt rates for both models, with clear dependencies on scale, terrain roughness and variations of temperature thresholds, among other quantities. The sub-grid parameterizations tested in this article are found to effectively compensate these effects at little additional computational cost.