CorridorKey for Resolve (Windows)

An OFX plugin that brings CorridorKey — Corridor Digital's AI-powered green screen keyer — into DaVinci Resolve 20 as a native plugin.

Note: This is a Windows-only release. The plugin uses Windows-specific APIs (named pipes, shared memory via CreateFileMapping) for IPC between the C++ plugin and Python backend.

CorridorKey produces physically accurate foreground color unmixing with clean linear alpha channels. This plugin wraps the full inference engine, delivering a native Resolve experience backed by PyTorch on your GPU.

Architecture

A thin C++ OFX plugin handles Resolve integration (UI, frame I/O, node graph), while a background Python process runs CorridorKey's inference engine via PyTorch. Communication uses Windows shared memory for frame data and named pipes for control messages.

DaVinci Resolve  <-->  OFX Plugin (C++ DLL)  <-->  Python Backend (PyTorch)
                       shared memory + pipes         CorridorKey + BiRefNet

The backend auto-launches when you first apply the plugin in Resolve — no manual startup required.

Requirements

• Windows 10/11
• DaVinci Resolve 20 (Free or Studio)
• NVIDIA GPU with CUDA support (tested on RTX 4090)
• Python 3.13 (3.14+ lacks PyTorch wheels)
• Visual Studio 2022 Build Tools (only if building from source)

Quick Install

1. Install Python 3.13 — check "Add Python to PATH" during setup
2. Download the latest release from GitHub Releases
3. Extract the zip
4. Double-click install.bat

The installer handles everything automatically: copies the OFX plugin, creates a Python environment, installs PyTorch + CUDA + Triton, clones CorridorKey, and downloads model weights.

Building from Source

If you prefer to build the C++ plugin yourself instead of using the pre-built release:

cd plugin
cmake -B build -G "Visual Studio 17 2022"
cmake --build build --config Release

Then run install.bat or python backend/install.py to set up the backend.

Dependencies Installed

Package	Purpose
torch + torchvision (CUDA 12.4)	Neural network inference
triton-windows	torch.compile kernel generation (critical for performance)
transformers	BiRefNet model loading via HuggingFace
timm	Vision model backbones
opencv-python	Image processing (resize, morphology)
numpy	Array operations

Usage

Color Page

1. Open DaVinci Resolve and go to the Color page
2. Add a corrector node and apply CorridorKey from the OFX plugins list
3. Connect your green screen plate to the Source input
4. The plugin auto-launches the backend on first use

Fusion Page

1. Go to the Fusion page
2. Add a CorridorKey node from the OFX tools
3. Connect your MediaIn to the Source input
4. Optionally connect an external mask to the AlphaHint input

Input Pins (Color Page)

Pin	Color	Purpose
Source RGB	Green	Green screen footage (required)
Source Alpha	Blue	Alpha channel of source
Alpha Hint	Blue	External mask input for External mode (optional)

First-Time Startup

On the first render after launching the backend, there is a one-time warmup delay of ~60 seconds. This is torch.compile generating optimized CUDA kernels via Triton for your specific GPU. Subsequent frames render in ~1.5-2 seconds. This compilation is cached by PyTorch and will be faster on future launches.

Parameters

Parameter	Type	Default	Description
Mode	Choice	Auto	Alpha hint source: Auto (BiRefNet) or External (user-provided mask)
BiRefNet Model	Choice	General	BiRefNet variant: General, Portrait, or Matting
Input Colorspace	Choice	sRGB	Colorspace of source footage: sRGB Gamma or Linear
Despill Strength	Double	1.0	Green spill removal intensity (0–10)
Auto Despeckle	Boolean	On	Remove small isolated alpha artifacts
Despeckle Size	Integer	400	Minimum pixel area threshold for despeckle
Refiner Strength	Double	1.0	CNN refiner multiplier (0–5)
Output Mode	Choice	Processed	What to output downstream

Output Modes

Mode	Description
Processed	Premultiplied linear RGBA — primary output for compositing
Matte	Alpha channel visualized as grayscale
Foreground	Straight (unpremultiplied) foreground with alpha
Composite	Preview composite over checkerboard

Input Colorspace

Set this to match your Resolve project's working colorspace:

• sRGB Gamma: Standard footage, Rec.709 projects
• Linear: ACES, linear workflow, or EXR footage

All outputs are delivered in the matching colorspace so they integrate correctly into Resolve's pipeline.

Performance

Tested on RTX 4090 with 2880px footage:

Stage	Time
BiRefNet alpha hint	~1.5s (cached per frame)
CorridorKey inference	~1.5s (fp16 + torch.compile)
Post-processing	~50ms
Total (first frame)	~3s
Changing output mode/despill	~50ms (cached)

Performance Features

• fp16 inference: Both CorridorKey and BiRefNet run in half precision for maximum GPU throughput
• torch.compile + Triton: Generates optimized CUDA kernels, ~3x faster than eager mode
• GPU resize: Upscaling from model resolution back to frame resolution happens on GPU (bicubic)
• Inference caching: Changing output mode, despill, or despeckle doesn't re-run the neural network
• BiRefNet caching: Alpha hint is cached per-frame, not regenerated on parameter tweaks
• Auto-reconnect: Plugin automatically reconnects if the backend restarts

File Locations

What	Path
OFX Plugin	C:\Program Files\Common Files\OFX\Plugins\CorridorKeyForResolve.ofx.bundle\
Backend install	%APPDATA%\CorridorKeyForResolve\
Virtual environment	%APPDATA%\CorridorKeyForResolve\venv\
CorridorKey repo	%APPDATA%\CorridorKeyForResolve\CorridorKey\
Model weights	%APPDATA%\CorridorKeyForResolve\CorridorKey\CorridorKeyModule\checkpoints\

Troubleshooting

Plugin shows passthrough (no effect)

The backend isn't running or the connection was lost. The plugin auto-launches the backend, but if it fails:

%APPDATA%\CorridorKeyForResolve\venv\Scripts\python.exe %APPDATA%\CorridorKeyForResolve\server.py

Node turns red in Fusion

The project may not be set to 32-bit float processing. Check Project Settings > Color Management.

"torch.compile failed" in backend logs

Install or update Triton:

%APPDATA%\CorridorKeyForResolve\venv\Scripts\python.exe -m pip install triton-windows

Ensure the Triton version is compatible with your PyTorch version.

Slow first frame (~60s)

This is normal — torch.compile is generating optimized CUDA kernels for your GPU. This only happens once per backend launch and is cached by PyTorch.

IPC Protocol

Communication between the C++ plugin and Python backend:

• Named pipe: \\.\pipe\CorridorKeyForResolve (JSON control messages)
• Shared memory: CorridorKeyForResolve_Input, _Output, _AlphaHint (float32 RGBA frames)
• Frame header: 16 bytes (width u32 + height u32 + channels u32 + reserved u32)
• Max frame size: 4096×4096 (configurable)

License

CC BY-NC-SA 4.0 — matching the upstream CorridorKey license.