[Inference.Core] AIO Aux Preprocessor

This is the "just let me try a few things" node. Instead of dragging in a separate Canny node, a separate HED node, a separate depth node just to see which one looks best on your image, you drop one AIO node and flip through a dropdown of over 40 preprocessors - Canny, HED, BAE normal maps, scribble variants, luminance, and more. It's the fastest way to A/B different ControlNet conditioning types without rewiring your graph every time. This copy comes from LykosAI, the Stability Matrix team, whose Inference_Core_ class prefix marks it as their fork of the well-known comfyui_controlnet_aux collection - same idea, repackaged so it doesn't stomp on the original pack's node names if both are installed.

What it is, and the catch

The tradeoff is honest and worth knowing up front: AIO gives you fast access to a preprocessor, but not that preprocessor's own tuning knobs. Pick a Canny-style option from the dropdown and you get it at its baked-in defaults - no low/high threshold sliders. Pick a depth model and you don't get to choose the checkpoint size. So the sane workflow is: use AIO to quickly find the preprocessor that gets you closest to what you want, then swap in that preprocessor's dedicated standalone node (this pack ships one for nearly everything AIO can do) once you actually need to fine-tune it.

The inputs and outputs that matter

• image - required, your source image.
• preprocessor - the whole point of the node. Defaults to none (a passthrough, occasionally handy as a bypass switch). The dropdown lists 43 options spanning line extractors, depth/normal estimators, segmentors, and more.
• resolution (default 512) - the working resolution for whichever preprocessor you selected.

One output: IMAGE, the hint map, which wires straight into a ControlNet Apply node. One thing to flag if you're coming from this pack's dedicated pose nodes: AIO only outputs an image, never a POSE_KEYPOINT. If you pick a pose preprocessor from the dropdown you get the rendered skeleton picture but not the structured keypoint data - for that you need the standalone DWPose Estimator or OpenPose Pose node instead.

How to install it

ComfyUI Manager: search ComfyUI-Inference-Core-Nodes, install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes

then run install.py, or pip install -e .[cuda12] (swap the extra for cuda, rocm, directml, or cpu to match your setup - this is what gets you a working accelerated onnxruntime rather than a mismatched one). Restart ComfyUI. Because AIO can invoke any of the pack's preprocessors, model weights for whichever one you pick download from Hugging Face the first time you actually use it - so switching to a new option in the dropdown for the first time means a pause while it fetches files.

Common issues & troubleshooting

If a preprocessor you expect just isn't in the dropdown, a dependency for it likely failed to install - this pack, like its upstream, skips nodes it can't import rather than crashing the whole thing, so check the ComfyUI console at startup for an import error tied to that specific preprocessor.

If pose-family or detector-backed options (anything touching .onnx model files under the hood) run painfully slowly, that's the same onnxruntime/CUDA-version mismatch that plagues this pack's standalone DWPose and AnimalPose nodes - the model silently falls back to CPU and prints a warning about missing acceleration providers in the console. Match your onnxruntime-gpu install to your CUDA version, or just reach for the standalone node instead, since several of them default to TorchScript checkpoints that sidestep onnxruntime entirely.

And if the pack itself fails to load - ComfyUI Manager flags it "conflicted," or you get a bare ModuleNotFoundError on startup - that's most often a collision with another controlnet-preprocessor pack sharing dependencies or class names, not a broken AIO node specifically. Worth ruling out before troubleshooting the dropdown itself.