The head image changes to pixel wind in seconds. Uncle homestead teaches AI by himself to create master level, which is popular on twitter

Recently, a Japanese uncle Sato created a AI portrait website AI gahaku. In 10 days, the number of user visits increased from 0 to < / P > < p > Sato suffered from Asperger’s disease, also known as “social phobia”, a real genius disease. Generally speaking, such patients will have special potential in a certain field, but they are not suitable for ordinary people’s lives. < / P > < p > this face portrait generation technology is based on the pix2pix model of cgan (conditional generation confrontation network) built in tensorflow. < / P > < p > different from Gan, cgan emphasizes playing in specific scenes. For example, AI gahaku can only aim at faces. If the user uploads cat face, it cannot generate results. The characteristic of Gan lies in the self calculation and automatic update of loss function, which makes Gan itself have the potential to combine with transfer learning. In recent years, Gan is more used as a means to achieve the goal of transfer learning. Transfer learning provides specific direction for the application of GaN. < / P > < p > tensorflow’s built-in pix2pix is a cgan based image to image translation model. Since the model itself is mature, it can be applied to black and white image coloring, image style change and other scenes. < / P > < p > specific to the use of uncle, using pix2pix model, uncle actually limits the model to the generation of master portraits, which also reflects the strong expansibility of the model itself. < / P > < p > in pixel me, another work of Sato, the pix2pix technology is also used, but the 8-bit pixel style head portrait is generated, which is a model with multiple uses. Of course, the actual effect varies from person to person. After all, strictly speaking, uncle is redeveloping the existing model, and the original algorithm and data structure have not changed, so the performance and effect are limited by the original model. Uncle’s life experience can be said to be complex. After dropping out of University, he worked as a baker and participated in nursing school training courses, but he felt that they were not suitable for him, until he decided to use his intelligence to devote himself to AI. < / P > < p > with the help of Google lab’s computing resources, uncle learned from scratch the tensorflow tutorial. However, uncle is really suitable for AI instead of steaming cakes. < / P > < p > in the process of learning, Sato found his own direction, transfer learning and Gan, can better achieve image regeneration, using different image training data sets, can learn and generate specific scenes. < / P > < p > Gan is a typical unsupervised learning method. Its core principle is to let two neural networks “confront” and obtain the optimal result by continuously optimizing the parameters. < / P > < p > thanks to the generosity of Google, the website servers and computing resources used by uncle are relatively cheap, which is about $20 a day. Uncle said that he would not seek commercialization in the short term within the scope of acceptable ability. Video Number assistant internal test online! Four functions let you send 1g video on the computer