What it does

PI-REC is a conditional GAN that takes extreme sparse inputs—think binary edge maps and flat color blobs—and progressively reconstructs detailed images. The repo ships both a batch command-line mode and a GUI drawing tool where you scribble edges, drop color hints, and watch the network fill in the rest.

The interesting bit

The “progressive” angle is the hook: the network is trained to handle reconstruction in stages, edge domain first, color domain second, rather than demanding a complete sketch. The authors are upfront that this is image reconstruction research, not a commercial auto-painting product—though the demo GIFs of anime-style face generation have clearly attracted attention.

Key highlights

• Reconstructs from “extreme sparse inputs” (the authors’ phrase), not just clean line art
• Interactive GUI mode for real-time human-machine creation
• Pre-trained models available for CelebA and anime-style faces
• CPU inference is possible by editing config.yml, though an NVIDIA GPU is expected
• Academic paper with full architecture diagrams; not a black-box repo

Caveats

• Training code is explicitly incomplete: “The codes of training for release are no completed yet, also waiting for release license of lab”
• Baidu Netdisk weights were taken down in April 2021; Google Drive is the only remaining source
• The authors warn it is “far from reaching the level of civilian or professional auxiliary drawing”
• Code structure is modified from prior work (Anime-InPainting / Edge-Connect), so this builds on established foundations rather than reinventing them

Verdict

Worth a look for researchers in conditional image generation or anyone curious about edge-to-image translation with a hands-on demo. Skip it if you need trainable code or production-ready art tools.