CorridorKey is a neural network tool that solves green screen keying by unmixing foreground and background colors at the pixel level.
The problem it addresses is that traditional green screen keying produces harsh, unusable edges because pixels at the boundary between subject and background contain a blend of both colors. CorridorKey takes a different approach: instead of creating a binary mask, it uses a neural network to reconstruct the true, unmultiplied color of the foreground object for every pixel, including semi-transparent ones like motion blur and out-of-focus edges. The model outputs both a clean linear alpha channel and the separated foreground color, effectively solving what it calls the unmixing problem rather than just guessing opacity levels.
The tool works with both green and blue screens, automatically detecting which one is in use or accepting manual specification. It handles 4K resolution by dynamically scaling inference while using a native 2048x2048 backbone for high fidelity. Output is in industry-standard 16-bit and 32-bit linear float EXR format for direct integration into compositing software like Nuke, Fusion, or Resolve. A morphological cleanup system can automatically remove tracking markers or small background artifacts. The project notes that an artist-friendly user interface version exists as a separate package for those preferring easier installation.
The project is brand new and explicitly invites community contribution to identify and patch issues. Development is active with the maintainer soliciting feedback through a dedicated Discord community and encouraging pull requests. The codebase uses uv for dependency management to streamline installation across operating systems. The maintainer acknowledges that not everything has been tested and frames the release as an open invitation for collaborative improvement rather than a finished product.
