← all repositories
nikopueringer/CorridorKey2.0

The green beast has been slain, says the README, and it's not wrong

Corridor Digital's ML keyer replaces green-screen spill wrestling with a PyTorch model that pulls mattes from ordinary footage — with or without hints.

Collecting fresh signals — velocity needs a few days of history.
collecting data…
star history

What it does CorridorKey 2.0 is a machine-learning matting model that separates foreground subjects from background in video — no green or blue screen required. It runs two stages: stage 1 processes overlapping temporal windows (Hann windows) to build context, stage 2 streams full-resolution predictions from those combined hidden states. It builds on three frozen foundation models — NVIDIA’s C-RADIOv4-SO400M, Microsoft’s MoGe-2, and Google DeepMind’s RVM small — with weights fetched from Hugging Face. Native resolution is always preserved; the model pads internally and crops back.

The interesting bit The hint system is the clever part. You can run it blind (model picks the subject), give it a single first-frame matte, or feed a full hint video — and a --hint_quality value tells the model how much to trust each hint, from 0.95 for artist-quality mattes down to 0.5 for loose scribbles. Carried hints let the model’s own overlapping-window predictions feed back as guidance, trading control for temporal stability.

Key highlights

  • No-hint, first-frame-hint, and full-hint-video inference modes with tunable hint trust
  • Automatic Hann window sizing based on free VRAM and clip resolution (80/40 down to 1/1)
  • HDR EXR output in scene-linear RGB, preserving values above 1.0 and negatives
  • Despill strength control and optional linear-light cutout for VFX pipelines
  • Free for commercial and personal use; a paid plugin for common media software exists separately

Caveats

  • Requires an NVIDIA RTX 30-series or newer GPU; 4 GB VRAM for 1440p, 6 GB for 4K, plus ~5 GB of weights
  • No-hint mode can pick a different subject than you intended; full hints can inherit the hint’s systematic bias
  • 96/48 windows were unstable at native 4K even on a 96 GB card, and 80/40 used most of that card’s VRAM

Verdict VFX artists and Python tinkerers who want ML keying without a chroma setup should absolutely try this. If you’re on AMD/Apple silicon or want a turnkey tool, the paid plugin (or something else) is your route.

Frequently asked

What is nikopueringer/CorridorKey2.0?
Corridor Digital's ML keyer replaces green-screen spill wrestling with a PyTorch model that pulls mattes from ordinary footage — with or without hints.
Is CorridorKey2.0 open source?
Yes — nikopueringer/CorridorKey2.0 is an open-source project tracked on heatdrop.
What language is CorridorKey2.0 written in?
nikopueringer/CorridorKey2.0 is primarily written in Python.
How popular is CorridorKey2.0?
nikopueringer/CorridorKey2.0 has 1k stars on GitHub.
Where can I find CorridorKey2.0?
nikopueringer/CorridorKey2.0 is on GitHub at https://github.com/nikopueringer/CorridorKey2.0.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.