CorridorKey 2.0: Get a Clean Green Screen Key From One Painted Frame, Free Even for Client Work
The model is free and you may use it on paid jobs. The trick that makes it reliable is telling it, once, who the subject is.
Every green screen key fails in the same places. Flyaway hair. A hand moving fast enough to blur. A glass of water, a veil, cigarette smoke. A traditional keyer looks at colour, sees something halfway between green and skin, and has to guess. The result is the familiar green fringe, or a matte that shimmers from frame to frame as the guess changes.
CorridorKey 2.0, the neural keyer from Corridor Digital, attacks those exact places. Instead of thresholding colour, it predicts two things for every frame: a matte (the black-and-white image that says how see-through each pixel should be) and the true colour of the foreground with the green taken out. That second output is why hair and blur come through without a fringe.
The surprising part is the licence. You can download the model for free, and its terms say: "Use CorridorKey for whatever you'd like, including processing footage for commercial projects." Free AI models usually come with a non-commercial clause. This one does not, for most people.
But a smart keyer still has to know what it is keying. That is where one painted frame changes everything.
Why the model needs a hint
Run CorridorKey with no help and, in the README's words, "The model decides what the foreground subject is." Most of the time it gets it right. On a busy plate it may not: "The model may make a different subject-selection decision than you intended." Two actors, an actor holding a prop, a stand partly in frame. The model has to pick, and its pick may not be yours.
The fix is a hint. CorridorKey accepts three levels of help:
- No hint. Fastest to set up, fine for one clear subject.
- A first-frame hint. One black-and-white image of frame one. White is the subject, black is the background.
- A full hint video. A rough matte for every frame, for example from a roto pass or another automatic tool.
The first-frame hint is the sweet spot for most jobs. You paint one frame, which takes a few minutes, and the model carries your intention through the clip.
How it carries one frame across a whole shot
Two ideas make this work, and neither needs maths to understand.
First, the model looks at overlapping chunks of frames rather than one frame at a time. Each frame gets judged with its neighbours in view, which is what stops edges from crawling. Corridor's product page calls this "time awareness," and it is the main thing 2.0 adds.
Second, it feeds its own answers forward. Once it has keyed the early frames using your hint, it uses those results as the hint for the frames that overlap them. The README calls this "carry," and it is on by default when you give a first-frame hint. Your painting starts a chain, and the model keeps passing it along.
That chain is also where things go wrong, which we will get to.
What you need
For the free model: an NVIDIA graphics card, "currently RTX 30-series or newer." CorridorKey says 4 GB of graphics memory is enough for footage up to 1440p and 6 GB for 4K. It checks how much memory your card has free and picks the chunk size for you. On small cards, the README says, it drops "to windows of only a few frames, or a single frame," and results "remain good, with somewhat less frame-to-frame consistency." You also need ffmpeg, the free video tool, installed.
On a Mac: the official free version is NVIDIA-only. On October 21, Peter Teravskis published an Apple Silicon version of the same trained model, covered at the end of this piece.
If you would rather not touch a terminal: Corridor sells a plugin for DaVinci Resolve, Nuke, After Effects, Premiere Pro and Blender. It is $49.50 for a personal licence during the beta, rising to $99 on November 21, as a one-time purchase. Mac support for the plugin is in beta. The routine below works the same way in principle, but this article follows the free version.
Put this into practice
Here is the routine, start to finish, for a medium-length green screen shot with one actor and some hair to worry about.
1. Install it once. From the CorridorKey 2.0 repository, in a Python virtual environment (a self-contained folder that keeps these installs away from the rest of your system):
pip install torch==2.9.1 torchvision==0.24.1 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt
python scripts/download_weights.py
The download is about 5 GB, because it also fetches three helper models from NVIDIA, Microsoft and Google DeepMind that CorridorKey builds on.
2. Paint the hint. Export frame one of your clip. In Photoshop, Affinity or anything with a brush, paint the actor solid white and everything else solid black. Do not chase individual hairs. You are telling the model who, not drawing the matte for it.
3. Decide how much to trust your painting. The --hint_quality setting tells the model how much weight to give your hint. The README's guide: 0.95 for a "high-quality matte or artist-provided hint," 0.75 for a "rough but useful automated hint," and 0.50 or below for a loose scribble. A quick brush job sits around 0.75.
4. Test on a slice, not the whole clip. Use --num_frames to key a short run, and --start_mode middle to take it from the middle of the shot, where trouble usually lives. A full render is the last step, not the first.
5. Ask for a checking output. Add checker to your outputs. It lays the keyed actor over a checkerboard with the green spill removed, which shows edge problems at a glance. Adjust --despill_strength (0.0 to 1.0, default 0.5) if green is still creeping onto skin or hair.
6. Render for the comp, not the preview. For final delivery, use --output_format exr. EXR is a file format compositors use for high-quality image sequences, and here it keeps brightness values above normal white, so bright highlights do not get flattened. If your plates are already linear EXR files, add --frame_dir_linear. Normal video outputs are, per the README, "clipped to [0, 1]," which means highlights above white are lost.
The lazy version. On Windows or Linux, drag clips onto the CorridorKey_DRAG_CLIPS_HERE launcher. If a file named <clip>FrameHint.png sits next to the clip, it picks up your painted frame automatically. Name your hints that way from the start and batch jobs get easy.
Where it breaks
Long, busy shots. The README is direct about the first-frame hint: it "can drift if the shot is long, crowded, or has large occlusions," meaning things passing in front of the subject. The carry chain that makes it convenient also carries mistakes forward. If someone walks in front of your actor, cut the shot into sections and paint a new first frame for each one.
A bad full hint steers a bad key. If you feed a full hint video, the README warns the output "will tend to follow systematic bias in the hint," so a matte that is too tight, too loose or on the wrong person pulls the result that way. A rough automatic matte deserves a lower trust setting.
Log and camera colour spaces. It "does not automatically convert ACEScg, Rec.2020, LogC, S-Log, V-Log, PQ, HLG, or camera-raw color spaces." In plain terms: if you shoot log or a camera's flat profile, convert the footage to normal display colour (or linear) first, or the key will misjudge your plate.
Not every control is finished. The README says its brightness and contrast controls are "useful for controlled tests, not for ordinary production use." Leave them alone on client work.
The licence has edges. You can use your output commercially. You may not sell the tool or charge for access to it, and companies with more than US$20 million in annual gross revenue need to buy a plugin seat for each user.
No free web demo. You have to install it or buy the plugin to try it.
The Mac route
The Apple Silicon version lives at corridorkey-v2-mlx, with the model files at cdelta/CorridorKey-v2-MLX. The port's author checked that every stored set of numbers in the model (738 of them) matches the official release exactly; only the software that runs it changed. On his M4 Pro with 48 GiB of memory, 32 frames of 720p footage took about 37 seconds at full quality, using about 7 GiB. You need Python 3.12, and the footage has to fit one of its prebuilt frame sizes (720p runs as 1280×768 and is cropped back).
Two cautions. It has been tested on that one machine only, and the author says the memory figures are not a minimum. And it currently handles normal display footage only, without the EXR high-brightness workflow or audio. The smaller, lower-precision 8-bit version saves memory but changed the matte by up to 6.9 percentage points on the author's glass test, so stick to full quality for hero shots.
The bigger point
For years, the best keys came from expensive plugins or patient roto artists. CorridorKey does not replace the patient artist. It changes what their time goes to. Instead of fighting a fringe on every hair, you spend five minutes telling the model who the subject is and the rest of your time on the shots that drift.
Pick one shot from your last job that you remember fighting. Paint frame one, key a 48-frame slice from the middle, and look at the checkerboard. You will know in ten minutes whether this belongs in your pipeline.