ControlNet Preprocess Canny

ControlNet is how you tell a diffusion model where things go, not just what to draw, and Canny edge maps are its oldest, most reliable language: hard, thin lines that say "the structure of the image goes here." This node (Runware_controlnet_preprocess_canny) extracts exactly that from an IMAGE - one input, one output, and the thresholds you'd expect from the classic preprocessor.

It's part of the Runware pack's builder pipeline, so it's usually not a standalone step. The typical graph is: load a reference image → this node produces the edge map → feed it into the pack's ControlNet builder → wire that into a model node's controlNet socket (like RunwareArch_sdxl). That's the whole ControlNet recipe, and this node is the first ingredient. If you've used ComfyUI's built-in Canny preprocessor locally, this is the same operation, just running on Runware's side so it plugs into their cloud models.

How it works

The node sends a controlNetPreprocess request (runware:controlnet-preprocess@canny AIR) with your image base64-encoded, and the edge map comes back as an IMAGE. Canny detection works by finding sharp intensity gradients and keeping only the ones that connect into contours - two thresholds decide how much survives:

• settings.highThresholdCanny - default 200. Edges with gradient strength above this are kept unconditionally. Raise it for fewer, cleaner lines; lower it to catch faint structure.
• settings.lowThresholdCanny - default 100. Gradient strength below this is discarded; anything between the two thresholds is kept only if it connects to a strong edge.

The classic beginner move is cranking both up because the map looks noisy - but a too-thin edge map gives the downstream model almost nothing to anchor on, which shows up as weak structure adherence. The defaults are well-chosen; nudge, don't panic.

The inputs that matter

• image (required) - the reference IMAGE.
• settings.highThresholdCanny / settings.lowThresholdCanny - the two thresholds above. These are the only parameters that matter.
• ttl / outputFormat - the standard plumbing; irrelevant for the tensor output you'll actually wire onward.

The output is a single image - the edge map. Don't expect a pretty picture; it's a gray line drawing, and that's correct.

Installing it

Install the pack. ComfyUI Manager → search Runware → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Key from runware.ai/api-keys → ComfyUI Settings → Runware API key (or RUNWARE_API_KEY).

Where people get burned

• The edge map alone does nothing. This is a preprocessing node - it must feed a ControlNet builder, which feeds a model. Wiring it to a preview and wondering why nothing changed is the classic miss.
• Canny isn't for every subject. It's best for architecture, mechanical objects, and scenes with clear contours. For organic subjects (people, animals), the pack's other preprocessors - depth, softedge, openpose - are usually a better fit. The controlnet KB calls this out explicitly.
• Threshold extremes both bite. Too low = noise that confuses the model; too high = no structure to follow. Stay near the 100/200 defaults and tune in small moves.
• Cost and latency. Each preprocess is a metered cloud call; it's cheap but it's a round-trip. Local ComfyUI has this preprocessor built in for free - this node exists to serve the cloud model path.

Runware_controlnet_preprocess_canny is a small, correct tool: feed it a reference, get an edge map, feed that into ControlNet, get structure. Learn the two thresholds and it'll never surprise you.