With the development of game technology and artistic ability, the improvement of computing power of game platform, and the increasing demand of players for roles and personalization, "face pinching" has become a very common part of all kinds of games, especially MMORPG. Before officially entering the game, many players will spend hours "pinching" the character image they want, and having a favorite character image will also affect the player's retention to some extent. However, when players are not satisfied with their face-kneading works, but want to enter the game as soon as possible, or want to quickly establish the specific character image they want, what can be done to help players? Nowadays, with the help of deep learning technology, we can help players achieve this goal in the blink of an eye.

First of all, analyze this problem. We hope to establish such a system, whose input is the photos containing faces uploaded by users, and the output is the parameters of face pinching, including continuous skeletal parameters (displacement, size, rotation angle, such as eye size, eyebrow height, cheek fat, etc. ) and discrete maps and color numbers (such as eyebrow shape, lip color, beard style, etc.). ). So this is a regression problem. Our network simulates a mapping function from face to parameters. For face detection and recognition, we don't want to extract features from the training network, but choose a mature scheme, which uses faceboxes for face detection and facenet for face coding, thus simplifying the problem into a mapping function from face coding vector (5 12× 1) to face pinching parameter vector (here set as 46× 1).