CorridorKey BRP ONNX model conversions
Converted model graphs and unchanged trained weights for native inference. This free collection contains no application source, executables, native runtimes, user footage, projects, reference tensors, or credentials. No publisher endorsement is claimed.
CorridorKey 2.0 uses one spatially dynamic pack: native source sizes are reflect-padded to multiples of 128 and cropped after inference. The adaptive pack also has a shared dynamic temporal axis for Stage1, with bounded encoder/MLP/attention execution; the runtime contracts distinguish supported bounds from actually validated lengths. RVM, C-RADIOv4-SO400M and MoGe are foundation components. SAM is SAM2.1 Hiera SMALL video propagation. TransNetV2 detects scenes.
Original trained weights are preserved. Conversion changes include dynamic shape graphs, incorporated feature-normalization statistics, and ONNX layouts for native execution. The adaptive pack contains FP32 reference graphs and separately declared DirectML mixed-FP16 compute graphs sharing the same original FP32 weight data. Sensitive normalization/position/explicit Softmax operations and public I/O remain FP32; unreachable conversion nodes are removed. Mixed precision has numerical differences and does not inherit the FP32 reference profile's strict qualification. These conversions are not official upstream releases. Windows DirectML was validated; no claim is made that every ONNX provider or machine supports these graphs.
Original licenses and attribution
Each component retains its own terms. CorridorKey portions and derivative conversion contributions retain the complete Corridor Key Licence 2.0, including all additional terms. This is not a relicensing of the collection. See model-notices/NOTICE.txt and model-notices/MODEL_LICENSES.json for publishers, exact licenses, modifications and provenance. Licensed by NVIDIA Corporation under the NVIDIA Open Model License.