GettingStarted.md

May 23, 2026 · View on GitHub

Preparing the model
denox accepts ONNX models as the network description. Trained pytorch or tensorflow models can be exported to ONNX. For example with pytorch

net = Net() # <- your pytorch model!
example_input = torch.ones(1, INPUT_CHANNELS_COUNT, 5, 5, dtype=torch.float16)
program = torch.onnx.export(
    net,
    (example_input,),
    dynamic_shapes={
        "input": {2: torch.export.Dim.DYNAMIC, 3: torch.export.Dim.DYNAMIC}
    },
    input_names=["input"],
    output_names=["output"],
)
program.save("net.onnx")

We advise giving inputs and outputs names, as we will later need those to refer to specific tensors. Additionally above, we compile the model with a dynamic input size, if this is not specified the model will only work for the shape of the example_input. Further if dynamic shapes are used denox expects models to be valid for all input sizes. For U-Net architectures with skip-connections this means that inputs must be explicitly aligned.

class UNetAlignment(nn.Module):
    def __init__(self, net):
        super(UNetAlignment, self).__init__()
        self.net = net

    def forward(self, input):
        alignment = self.net.alignment
        H, W = input.size(2), input.size(3)
        pad_w = (alignment - (W % alignment)) % alignment
        pad_h = (alignment - (H % alignment)) % alignment
        aligned = F.pad(input, (0, pad_w, 0, pad_h), mode="replicate")
        output = self.net(aligned)
        return output[:,:,:H,:W]
net = UNetAlignment(Net())

Compile

denox compile net.onnx \
    --type input=f16 output=f16 \
    --storage input=ssbo output=ssbo \
    --format input=hwc output=hwc \
    --shape input=H:W:C \
    --assume H=1080 W=1920 \
    --db /tmp/gpu.db \
    -o net.dnx

This produces a compiled model artifact net.dnx

See denox compile --help for more information.

Benchmarking & Testing
- To get quick profiling results, about inference latency and dispatch schedules.
```
denox bench net.dnx --spec H=1080 W=1920
```
- To test if compiled model contains the correct weights, the easiest way is to run a single inference on a PNG.
```
denox infer net.dnx -i input.png -o output.png
```
Integration within Renderers
The produced artefact net.dnx is a flatbuffer, which contains a list of dispatches, including SPIR-V binaries, descriptor-set bindings, push-constants and dispatch sizes. When executed sequentially those dispatches infer the model. denox intentionally does not provide a runtime instead it relies on engine developers to write this parsing step manually either within a asset pipeline at compiletime or at runtime.