Relative classification accuracy for BLENDHUNTER with respect to SEP on the realistic CFIS-like postage stamps, with σ_noise = 14.5. The points are taken from the average relative classification accuracy from the ten training noise realisations and the error bars from the standard deviation. For close blends (red dot-dashed line), BLENDHUNTER outperforms SEP for any level of noise in the training data.