Paint over anime scars with an OpenCV paintbrush

What it does This project wraps the Edge-Connect deep-learning inpainting model in an OpenCV-based interactive tool. You paint over damaged or masked regions of anime images with a brush; the backend fills them in using edge-aware generative adversarial networks. It also bundles training optimizations like auto-saving weights and a condensed training phase.

The interesting bit The value is in the glue: the author turned a research codebase into something you can actually operate with mouse clicks and hotkeys. There’s even an optional “edge window” mode where you manually sketch structure lines to guide the reconstruction — half automation, half artist control.

Key highlights

• Interactive inpainting via python tool_patch.py with brush controls ([/] for size, n/e for patch modes)
• Optional edge-editing window to manually guide the GAN’s structural understanding
• Training workflow improvements: auto-save/load weights, merged fast-training phase, CPU/GPU toggle in config
• Pre-trained weights available (anime face dataset), though the README warns these struggle with full-body shots
• CC BY-NC 4.0 license — non-commercial use only

Caveats

• The pre-trained model is tuned for anime faces, not whole characters; broader scenes need custom training
• Keys 0 and 1 are listed as “Todo” in the manual — unfinished UI features still visible in the source
• PyTorch 1.0 required; 0.4 explicitly breaks

Verdict Worth a look if you need to batch-restore or interactively repair anime-style artwork and want a GUI rather than notebook cells. Skip it if you need commercial licensing or production-grade generalization beyond the provided face-trained weights.