Picsart-AI-Research/MI-GAN

What it solves

MI-GAN is designed to perform high-quality image inpainting (filling in missing or unwanted parts of an image) specifically for mobile devices. It addresses the challenge of achieving state-of-the-art results while maintaining a small model size and fast inference speed, making it suitable for on-device processing.

How it works

MI-GAN serves as a lightweight baseline for inpainting. It is trained using a teacher-student framework where a larger, more complex model (Co-Mod-GAN) acts as the teacher to guide the training of the smaller MI-GAN student network. The project provides tools to train both the teacher and student models, as well as scripts to export the model to ONNX format for seamless integration into mobile applications.

Who it’s for

This project is for developers and researchers interested in deploying AI-powered image editing tools on mobile hardware, as well as those studying efficient GAN-based image synthesis.

Highlights

• Mobile-Optimized: One order of magnitude smaller and faster than recent state-of-the-art approaches while maintaining low FID (Fréchet Inception Distance).
• Versatile: Capable of producing plausible results on both complex scene images and face images (FFHQ).
• Developer-Friendly: Includes a script to convert the entire pipeline—including preprocessing and postprocessing—into a single ONNX file.
• Integration Ready: Compatible with tools like IOPaint for immediate testing.