Object image completion

One or more neural networks for generating complete depictions of objects based on their partial description are disclosed. An image 114 of a whole object is generated, based on an image 106 of a portion of the object, using an encoder 108 trained using training data 102 produced from the output of a decoder 112. The neural network may comprise a generative model framework, which can be a variational autoencoder, a generative adversarial network (GAN) or a normalising flow. The decoder can be trained on a dataset comprising images of complete objects and excluding images of partial entities. The decoder may output a complete version of an incomplete picture input into the decoder. The decoder parameters may remain unvaried while training the encoder (Fig. 6). Two images may be entered into the encoder, with the resulting output being the first image which is partially occluded by features from the second picture. An associated training technique is also described.