A distribution-free approach is proposed to estimate the fracture size distribution from a given trace length distribution, assuming that the fracture network can be represented by a Poisson-disc model. This approach directly works on the experimental distribution of the observed trace lengths (corrected from sampling biases) and does not need to choose a parametric model for the trace length or fracture diameter distributions. It is more robust than existing models and provides an unbiased estimate of the cumulative distribution function of the fracture diameters. Its simplicity of use, accuracy and versatility are illustrated through synthetic examples.