Call inspect_ml_environment, then create_ml_node. Use exact existing node
names as input_node and optional reference_node: the latter is ground truth
for CopyCat/BigCat or a mask for Deblur. CatFileCreator has no image inputs.
Settings are class-specific native knob names, never arbitrary scripts/buttons.
Training nodes need epochs, modelSize, cropSize, batchSizeType,
batchSize, checkpointInterval, imageInterval and useMultiResolution
configured intentionally. The tool uses CPU and starts no work during creation.
For conversion, configure a CatFileCreator with absolute torchScriptFile,
channelsIn/channelsOut lists such as ["rgba.red","rgba.green","rgba.blue"],
nonempty modelId and outputScale [1,1]. Invoke conversion with a new .cat
output. Create an Inference explicitly using its modelFile; native channels
and custom model state are resolved by Nuke, not parsed by this adapter.
For bounded training, configure paired image inputs and a maximum of 1000 epochs, then request a new output directory. The owned child uses the same executable, NukeX interactive licensing and a temporary script snapshot. It creates real checkpoints/contact sheets and retains its stdout/stderr and final receipt. Snapshot serialization restores original knob text, Root name and dirty state; multiple-context paths fail closed. Native execution may fail for license, model/format/channel or device reasons even after class discovery succeeds. The worker restores the source working directory, in-memory script name and native plug-in search order before execution, so relative media, Root-name expressions and already loaded Gizmos keep their source evaluation context. The original script is never saved by the worker. Only paths already trusted by the running host are reused; startup profiles and temporary files stay isolated.
Use Core job status/cancel; do not invent an ML-specific job id or assume a timeout rolled back already written checkpoints. Do not enable networked training, change global device preferences, or silently download model weights. Inference/Deblur results need a rendered-content check. Upscale must additionally show doubled dimensions; a node declaration or small preview alone is insufficient.
The native 16.1 knobs differ from parts of the user guide: halfPrecision,
useMultiResolution, CatFileCreator's non-suffixed scripting names and its
nonempty model ID requirement are
confirmed in a real NukeX 16.1 host. Extra training validation/augmentation knobs,
custom TorchScript user knobs and resume are not exposed by this bounded API.
Primary references: