Once your model is prepared, you can easily upload it to the Model Registry or download any model you have permission to access. This section provides a simple, step-by-step guide to both processes.
Upload Guidelines
It is assumed here that you already have the Model File(s) and Model Card prepared. If this is not the case, please refer to the Model Preparation guidelines.
For custom model conversion, prefer uploading an ONNX NN Archive instead of a raw ONNX file whenever possible.The archive carries preprocessing, tensor metadata, and optional heads metadata into the conversion flow, which greatly reduces manual setup and usually produces a converted archive that is ready to use in DepthAI.
Step by step instructions:
Open Models and click Add Model
Open Hub, navigate to the Models section, and click Add Model.
Choose the upload type
Select the model type you want to upload. Use YOLO for supported PyTorch YOLO weights and Custom for ONNX NN Archives, raw ONNX, OpenVINO IR, TensorFlow Lite, or already compiled RVC model artifacts.
Choose the target platform for YOLO
If you selected YOLO, choose the target RVC platform for the conversion.
Fill in model details and upload the files
Fill in the required descriptors, upload the model file(s), and confirm the action with Add or Export. Use the reference sections below for the exact meaning of the common metadata fields and the YOLO-specific parameters.
Add another version if needed
If you want to store another version of the same model, click Add Version and upload it as a separate version entry.
Copy the model version reference
After the upload is complete, use the Copy button next to the version you want to reference from DepthAI or other tooling.
Upload Type Reference
Upload type
Use it when
YOLO
Uploading supported PyTorch YOLO weights (.pt) that Hub can convert directly
Custom
Uploading ONNX NN Archives, raw ONNX, OpenVINO IR, TensorFlow Lite, or already compiled RVC model artifacts
Common metadata fields
Parameter
Meaning
Model Name
Name used to reference the model through Hub
Model Version
Version identifier such as 1.0.0
License
License that governs the use and distribution of the model
Public
Controls whether the model is visible to other Hub users outside your team
Short Description
Short summary shown under the model thumbnail
Model Image
Preview image illustrating the model
Tasks
Machine learning tasks handled by the model
YOLO-specific parameters
Parameter
Meaning
Model Instance Name
Name of the converted model instance
Version
YOLO version of the uploaded model
Class Names
List of output class names
Shape
Input shape of the model
Quantization Data
Dataset used to calibrate quantized conversion for RVC3 and RVC4
POT Target Device
POT target device for RVC3
Hub allows upload of multiple versions of the same model, for example when models were exported with different input shapes or trained on different datasets.
For public models, follow a consistent version naming scheme.The version name should not include the model name and should contain only the descriptors needed to distinguish variants.Separate descriptors with a single space, and keep the input shape (input_height x input_width) as the last descriptor.
Can't find a model you are looking for or have issues with uploading your model? Feel free to fill-in the model request form. We will do our best to support the most requested models.
Download Guidelines
To download a model from Hub, first open the desired public or private model and select the version you want to inspect.
Open the model version
Open Hub, navigate to Models, open the model, scroll to Model Versions, and click the desired version.
Base Model
Choose Base Model
To download the unconverted model, open the menu in the top right corner and choose Base Model.
Start the download
In the pop-up window, click the model name to start the download.
Converted Model
Open the converted instance
To download a converted model, click the name of the converted instance from the selected model version.
Download the converted archive
In the pop-up window, click the .tar.xz file to start the download.
