MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices
--model migan- GitHub: https://github.com/Picsart-AI-Research/MI-GAN (opens in a new tab)
MI-GAN can produce plausible results both on complex scene images as well as on face images. The bubble chart on the right shows the advantage of MI-GAN network over state-of-the-art approaches. The size of the bubble signifies the relative number of parameters of each approach and the number inside of the bubble shows the absolute number of model parameters. MI-GAN achieves a low FID, while being one order of magnitude smaller and faster than recent SOTA approaches.