Progressive Image Generation Model Based on Multi-level Features

Aiming at the demand of high-quality appearance product generation, a progressive image generation model based on multi-level features is proposed in this paper. The algorithm divides the image generation task into multiple stages, and each stage aims to generate the features of a specific abstraction level. Start from the high-level features, and then introduce the lower level features step by step with the training. This incremental generation algorithm enables the model to first focus on high-level abstract semantic information and large-scale structural information, and then gradually shift its attention to more and more fine scales without learning all scales at the same time. The design of the algorithm is inspired by the process of drawing the picture by the painter. The painter first outlines an approximate outline in the image area, and then depicts the detailed texture according to the color of the reference image, so that the whole picture looks more harmonious and consistent. This incremental generation method not only makes the generation process more orderly, but also at each stage in the later order, because the features to be generated have completed the generation of the previous abstract level, the generation task at the current stage is equivalent to being simplified, that is, the incremental generation strategy can use the knowledge learned from the previous subtask to simplify the learning of subsequent subtasks, Therefore, it is easier to train and get better results than the traditional one-step generation algorithm.