OpenAI's 'Glow' in which AI makes a change in expression and synthesis of two faces without a sense of incongruity

OpenAI , established as a nonprofit research institute on AI, unveils a technology " Glow " that allows you to synthesize two facial photos without discomfort and to add attributes such as age, beard and smile to facial photographs Did.

Glow: Generative Flow with Invertible 1 x 1 Convolutions
(PDF file) https://d4mucfpksywv.cloudfront.net/research-covers/glow/paper/glow.pdf

Glow: Better Reversible Generative Models
https://blog.openai.com/glow/

Developed by Mr. Diecelaic Kingma of OpenAI and Mr. Prahurra Daliwar, Glow is an image generation AI using a lossless 1 × 1 convolution (invertible 1 × 1 convolutions). Glow also features automatic labeling from learned images, and automatically learns the face photographs of 30,000 people prepared for demonstration, it automatically selects "smile", "age", "fine details", "blond hair" A label such as "beard" is set. The set label can be operated freely later and it is possible to change the newly input image.

For example, in the case of Mr. Jeffrey Hinton 's face photo set on INPUT on the upper left, you can see "Smiling", "Age", "Blonde Hair", "Beard" By setting the value to the maximum and setting "Narrow Eyes (fine)" to the minimum (state in which the eye can be widely opened), it is possible to change to the face picture of OUTPUT at the upper right. A little uncomfortable feeling remains with the beard being black, but it is a face of a man who is not strange even if it is actually, and it does not look like AI's own judgment made.

In addition, Glow can also synthesize two images and change the degree of composition. If you set the face picture of Neil Dograse · Tyson to LEFT INPUT, the face picture of Mr. Rashida · Jones to RIGHT INPUT, set the MIX of red frame to the center position and synthesize the two images evenly, it will be added to OUTPUT A natural man's face was created that is not strange even though it is actually.

Also, if the value of MIX is shifted to the left side, although the characteristics of Mr. Drogs · Tyson of LEFT INPUT will become stronger, the generated face picture will also be a natural face ......

Even if I put it to Mr. Jones (right) side, there was no discomfort.

OpenAI also publishes a movie synthesizing Mr. Dariwar's picture and images of nine famous people. Click the following image to play the movie.

The picture of Mr. Dariwar lined up in nine rows gradually changed ... ...

With almost no sense of discomfort, nine people turned into face photographs of each celebrity. If the hairstyle or facial features are too different, the image on the way seems to be amusing, but as a movie it did not feel a big incongruity.

Mr. Kingma and Mr. Daliwar also said that they do not like the image deletion in the middle, "Glow sometimes gives unnatural changes", and in order to make a more natural change, "use an autoregressive model or VAE I want to improve it. "

OpenAI publishes Glow on GitHub , and anyone can try it freely. In order to operate the Glow is, TensorFlow and Horovod is necessary, for such specific method of operation documents of Glow has been described.