Custom models

In theory, the Nx AI Manager can run any vision model. While it offers seamless support for image classification and object detection, it is also capable of running other model categories, such as:

• Image/instance segmentation

• Keypoint detection

• Vision-language models (VLMs)

• Custom model developed from scratch using PyTorch or Tensorflow

For these advanced model types, however, external pre-processing and post-processing are required to:

• Prepare and format the model inputs

• Interpret and process the model’s raw outputs

If you’re experiencing difficulties running a custom model during integration, please contact the support team for assistance.

Model output types

The following table shows how different model output types appear in the Nx Meta Client and what is required to use them:

Model type	Output format	What appears in Nx Client	Notes
Classification	Category + confidence score	Text overlay via event rule	Works out of the box
Object detection	Bounding boxes	Boxes drawn on video feed	Works out of the box
Object counting	Count events	Text overlay / event rule	Requires counting postprocessor
Segmentation / keypoints	Masks / keypoints	Requires external postprocessor	See External Postprocessing