Researchers from MIPT, the Higher School of Economics, and the Institute for System Programming of the Russian Academy of Sciences have developed an algorithm that improves image generation even with a lack of training examples. The technology helps artificial intelligence identify characteristic features of an object and transfer them more accurately to new images.
The problem is especially noticeable when a user uploads only one photo. Instead of learning the object's features, the neural network memorizes the background and random details.
Ask it to draw a cat in space, and it will depict the animal on the same couch, but with space drawn around it.
Scientists have changed the mathematical approach to noise processing, which generative models use to create images. The algorithm shifts the neural network's attention to weak noises that retain information about the object's subtle details.
Thanks to this, the model better learns the characteristic features of the object and is less dependent on random features of the original image. At the same time, the technology requires almost no additional computational resources.