Skip to main content
Back

Corridor Crew's CorridorKey Kills 30 Years of Green Screen Rotoscoping

Corridor Digital's neural chroma keyer reconstructs true foreground color under green pixels, preserving hair, motion blur and translucency in 4K EXR.

Repo, Full Access
  • Corridor Digital released CorridorKey, an open source neural chroma keyer with 13k+ GitHub stars.
  • Reconstructs true foreground color plus linear alpha per pixel, preserving hair, motion blur and translucency.
  • Trained on synthetic 3D scenes with known ground truth rather than scraped green screen footage.
  • Handles 4K via a 2048x2048 backbone, outputs 16 and 32-bit EXR for Nuke, Fusion, Resolve.
  • Runs on 6-8 GB VRAM across NVIDIA CUDA, AMD ROCm, Apple Silicon MLX, and CPU.
  • License is CC BY-NC-SA variant, free for commercial projects but not for repackaged sale or paid API services.

Green screen keying has been visual effects' dirty little chore for three decades. Commercial tools handle the easy 80 percent, then a human rotoscoper spends hours cleaning up hair, motion blur, and semi-transparent edges frame by frame. A new open source project from the YouTubers behind Corridor Crew wants to end that.

CorridorKey is a neural network chroma keyer built by Niko Pueringer that reframes the problem entirely. Rather than building a better binary mask, it treats keying as a color unmixing task. Trained on synthetic 3D data, the model predicts the true un-multiplied foreground color and a clean linear alpha channel for every pixel, including motion blur, out-of-focus edges, and translucent materials. The repo has crossed 13,000 stars, and a community ecosystem with Nuke and DaVinci Resolve plugins has already sprung up around it.

Unmixing, not masking

The core insight is worth pausing on. When you film against a green screen, pixels near the subject edge are a physical mixture of foreground light and green screen light. A traditional keyer, or even a modern AI roto tool, has to decide whether each pixel is in or out, which destroys the delicate semi-transparent gradients that make hair and glass look real. CorridorKey instead extracts clean mattes from footage with complex transparency, and outputs in formats that slot directly into professional compositing pipelines.

Concretely, the model takes two inputs and produces four outputs. Inputs are the raw green or blue screen plate (sRGB or linear gamma) plus a coarse alpha hint, a rough black-and-white mask that only needs to roughly isolate the subject. Outputs are:

  • /Matte - the raw linear alpha channel as an EXR
  • /FG - the straight, un-premultiplied foreground color