File size: 3,417 Bytes
8c5283d bcb9f8c 8c5283d c871d57 8c5283d c871d57 bcb9f8c c871d57 8c5283d c871d57 8c5283d c871d57 8c5283d c871d57 8c5283d c871d57 8c5283d c871d57 bcb9f8c c871d57 bcb9f8c 8c5283d c871d57 8c5283d bcb9f8c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 | ---
license: other
license_name: corridor-key-licence-2.0
license_link: LICENSE
pipeline_tag: image-segmentation
tags:
- video-matting
- green-screen
- blue-screen
- chroma-key
- vfx
---
# CorridorKey 2.0 — fork
Quantised weights for **[alphastrata/CorridorKey2.0](https://github.com/alphastrata/CorridorKey2.0)**,
a heavily modified fork of [CorridorKey 2.0](https://huggingface.co/nikopueringer/CorridorKey2.0)
by [Corridor Digital](https://www.corridordigital.com). Inference is pure Rust
and CUDA (`ck` and the `ck-gui` desktop app); no Python is involved. All credit
for the model goes to Corridor Digital; please support them at
[corridorkey.com](https://corridorkey.com).
## Fork changes
- Rust launcher and Slint GUI with a bespoke CUDA kernel suite, markedly faster
than the reference PyTorch engine, with multi-GPU scheduling.
- Prebuilt PTX from Pascal (`sm_61`) to Blackwell (`sm_120`).
- bf16 and fp16 quants: half the size at fp32 quality.
## Usage
```
git clone https://github.com/alphastrata/CorridorKey2.0
cd CorridorKey2.0
scripts/dev-setup.sh
just run shot.mp4 --weights weights/corridorkey_v2.bf16.safetensors
```
See the GitHub README for hints, HDR/EXR output and memory settings.
## Files
| File | Contents |
|---|---|
| `corridorkey_v2.safetensors` | Original fp32 weights (2.66 GB) |
| `corridorkey_v2.bf16.safetensors` | bf16 quant (1.34 GB); suits RTX 30-series and newer |
| `corridorkey_v2.fp16.safetensors` | fp16 quant (1.34 GB); suits RTX 20-series and older |
| `foundation_stats.safetensors` | Normalisation statistics for the foundation-model features |
Both quants store matrices at half precision and keep 1-D tensors (biases,
norms) in fp32.
## Requirements
- NVIDIA GPU, GTX 10-series or newer; 4 GB VRAM up to 1440p, 6 GB for 4K
- `ffmpeg`
## Foundation models
CorridorKey runs on features from three frozen foundation models. They are not
included in this repository; the download script fetches them from their
original publishers, under their own licenses:
| Model | Publisher | License |
|---|---|---|
| [C-RADIOv4-SO400M](https://huggingface.co/nvidia/C-RADIOv4-SO400M) | NVIDIA | NVIDIA Open Model License |
| [MoGe-2 ViT-B](https://huggingface.co/Ruicheng/moge-2-vitb-normal) | Microsoft | MIT |
| [RVM small](https://github.com/google-deepmind/representations4d) | Google DeepMind | CC-BY 4.0 |
## Training data
All training data for CorridorKey 2.0 was generated in-house at Corridor
Digital using 3D rendering.
## License
Use CorridorKey for whatever you'd like, including processing footage for
commercial projects. What you make with it is yours.
- You may not sell this model, or charge for access to it, and you may not
offer it as a paid API or inference service.
- You may build on the code and model, but what you release must be freely
accessible, stay under this same license, and keep the CorridorKey name.
Donationware and pay-what-you-want are fine.
- Organizations with more than US $20 million in annual revenue need to
purchase a seat of the CorridorKey plugin for each person using this model.
See [corridorkey.com](https://corridorkey.com).
- If you run a commercial software package and want to integrate CorridorKey,
email contact@corridordigital.com.
The full terms are in [LICENSE](LICENSE), a variation of
[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) with
additional terms. |