CorridorKey addresses the challenge of extracting clean foreground elements from green screen footage, particularly delicate semi-transparent edges often lost by traditional keyers. It offers VFX professionals and researchers a neural network solution for physically accurate unmixing, enabling realistic composites with preserved detail and reduced manual effort.

How It Works

The core innovation is a neural network performing physically accurate unmixing. For each pixel, it predicts the true, un-multiplied foreground color and a linear alpha channel, reconstructing the foreground as if the green screen were absent. This approach effectively handles fine details like hair and motion blur, surpassing traditional methods that struggle with semi-transparent pixels.

Quick Start & Requirements

Installation uses uv for Python/dependency management; Windows users run Install_CorridorKey_Windows.bat, Linux/Mac users install uv and run uv sync. The CorridorKey_v1.0.pth model must be downloaded manually; optional GVM/VideoMaMa modules require CLI downloads. A minimum of 24GB VRAM is essential (~22.7 GB for native inference), with higher VRAM recommended for optional modules. NVIDIA GPUs are standard; experimental MPS support exists for Mac. Cloud instances or secondary GPUs are advised.

Highlighted Details

• Physically Accurate Unmixing: Generates straight foreground color and linear alpha, preserving fine details like hair and motion blur.
• Resolution Independent: Scales inference dynamically to handle high-resolution plates (e.g., 4K) using a 2048x20