Most chroma‑key workflows struggle at edges because pixels become a blend of foreground and screen color; that’s where CorridorKey reframes the problem. Instead of producing a binary matte, it predicts the true, straight (unmultiplied) foreground color and a linear alpha per pixel, which preserves subtle semi‑transparent details like hair, motion blur, and defocus.
What Sets It Apart
- Physically oriented unmixing: the model reconstructs straight foreground color rather than only estimating opacity, so compositing retains correct color information where traditional mattes produce haloed or desaturated edges.
- VFX‑friendly outputs: natively reads/writes 16‑bit and 32‑bit linear float EXR passes (Matte, FG, Processed, Comp) so results integrate directly into Nuke/Fusion/Resolve without destructive conversions.
- Resolution independence with high‑fidelity backbone: inference scales to 4K plates while using a native 2048×2048 backbone to balance fidelity and memory; community optimizations target consumer GPUs.
- Modular AlphaHint workflow: integrates optional AlphaHint generators (GVM, VideoMaMa, BiRefNet) to automate or refine the coarse mask the model expects, improving results for challenging subjects.
Who It's For & Trade‑offs
Great fit if you work in VFX or video post and need high‑quality keys from real footage (hair, motion blur, transparent fabrics). It’s useful for compositors who want pixel‑accurate color + alpha passes instead of binary mattes. Look elsewhere if you need a lightweight, real‑time webcam keyer on low‑end hardware — the project targets high fidelity and can require substantial VRAM for full‑resolution inference (though community builds reduce this). Also note the repo includes heavy optional modules (GVM/VideoMaMa) whose licenses may be non‑commercial.
Where It Fits
Use CorridorKey when the compositing task demands preservation of subtle color/alpha interactions that traditional keyers or hard AI mattes destroy — for final comp passes, VFX delivery, or when producing archival EXR layers for grading and finishing.
How It Works (brief)
The pipeline takes an RGB green‑screen frame plus a coarse AlphaHint and uses a neural unmixing model to predict per‑pixel straight foreground color and a linear alpha. The repo provides orchestration, optional AlphaHint generators, Docker support, and device/back‑end paths (CUDA, ROCm, MPS, MLX) to run inference across platforms.
