--- 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.