New GAN Method for Full-Body Image Generation Proposed

Neural networks creating photorealistic portraits or even 3D scenes is not a new concept. However, AI can more or less reliably generate faces, but full-body images are still difficult due to the variation in poses, clothes, hairstyles, and perspective. 

InsetGAN introduces a method of combining multiple GANs where one GAN generates a global canvas (for example, a human body) and other GANs, or insets, focus on different parts (like faces or hands) that can be seamlessly inserted onto the canvas.

Combining a full body GAN with a high-quality face GAN helps InsetGAN produce "plausible-looking" humans. The method concentrates on parts that often contain artifacts, such as faces or shoes, and combines them with an appropriate body.

Here are some more results:

Check out the science behind the method here and don't forget to join our new Reddit page, our new Telegram channel, follow us on Instagram and Twitter, where we are sharing breakdowns, the latest news, awesome artworks, and more.