edge-connect

Why we included this project

For photo restoration, object removal, or filling damaged regions, EdgeConnect is a clean reference implementation of a two-stage adversarial inpainting method from an ICCV 2019 paper. Instead of predicting missing pixels directly, an edge generator first hallucinates the edges of the hole, then a completion network uses those edges as structure to paint in plausible content. The authors modeled the approach on how artists work, lines first, color next, and the result is better fine detail in filled regions. The repo ships with training and test scripts and works with standard datasets like Places2 and CelebA, with support for both regular and irregular masks, so it works as a starting point for researchers reproducing results or engineers prototyping content-aware fill. One caveat: it carries a non-commercial Creative Commons license, so check your intended use before building a product on it.