Tiddeman, BernardStirrat, MichaelPerrett, DavidLouise M. Lever and Mary McDerby2014-01-312014-01-3120053-905673-56-8https://doi.org/10.2312/LocalChapterEvents/TPCG/TPCGUK05/105-111The ability to combine multiple images to produce a composite that is representative of the set has applications in psychology research, medical imaging and entertainment. Current techniques using a combination of image warping and blending suffer from a lack of realism due to unrealistic or inappropriate textures in the output images. This paper describes a new method for improving the representation of textures when blending multiple facial images. We select the most likely value for each pixel, given the values of the neighbouring pixels, by learning from the corresponding values in the training set i.e. we use a Markov Random Field (MRF) texture model. We use a multi-scale neighbourhood and separate low and high frequency information using a wavelet transform. This ensures proper correlations of values across spatial scales and allows us to bias the global appearance to the mean for the set, while selecting more specific texture components at higher resolutions. We validate our results using perceptual testing that shows that the new prototypes improve realism over previous techniques.