Contact Us
List of Notch Block articles:
In Notch 2026.2, support was added for ONNX (Open Neural Network eXchange) format AI Models. This enables advanced image segmentation, background removal, feature tracking and more. Blocks featuring ONNX models will require careful setup and testing.
PIXERA and Notch do not provide support for creating, training, modifying, or troubleshooting ONNX models. It is not guaranteed that all ONNX models will be compatible with Notch, PIXERA, specific GPU architectures or driver versions.
Proceed at your own risk and test thoroughly before entering a show environment.
Warning: NVIDIA Virtual Background and ONNX AI models are not guaranteed to be compatible, so should not be contained in the same Notch Block.
Prerequisites
First, thoroughly read Notch's documentation on the AI model nodes and resources.
There are specific system requirements to run ONNX models in Notch Builder, as well as in Notch Blocks:
- NVIDIA CUDA
- TensorRT
- Recent NVIDIA Driver
The only officially supported PIXERA servers will be those with sufficiently up to date NVIDIA drivers, Currently, only those that have Blackwell cards.
Server Setup
The key steps from the Notch documentation are highlighted below. If there are multiple servers that will be required to render a Notch Block with an ONNX AI model, they will all need to be set up as shown. This includes Director and Backup servers, not just the main output server. Complete this list with reference to the Notch manual, as linked above.
- Install CUDA 12.9
- Set up the same file path for the .onnx and .trt files. This needs to be the same as on the machine where the Notch Block is built. For example ‘C:\tensorrt’ and ‘C:\onnx’.
- Set up the required environment variable.
- Download TensorRT and move the relevant DLLs.
Using Blocks with ONNX AI Models
The first time a Block using an ONNX model is loaded, the model needs to be unpacked and .trt files need to be generated. This initial process can take a considerable time. The .trt files are specific to the CUDA version, NVIDIA driver version and hardware running the block. Unless you're sure these are identical between machines, it is not advised to copy .trt files between servers. It is recommended to load the block on every machine to allow this process to occur uninterrupted before entering a show environment. Once the .trt files are present, block loading times should be much closer to normal PIXERA operation.
Once the initial loading and .trt generating process is complete, there should be no change to the usual workflow of interacting with Notch Blocks in your PIXERA projects.
PIXERA resource distribution will only transfer .dfxdll files (Notch Blocks) and will not transfer ONNX models or TRT files between machines.
Pixera 26.1 R 1 | 05/08/2026 | B.C.