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.