arkennemasis fal · Flux LoRA Canny · $0.035/MP

Canny conditioning is the blunt instrument of the ControlNet world and it's still the best one for a specific job: holding the shape of a reference while you replace the substance. A doorframe stays a doorframe, a wheel stays round, the roofline stays where it was - and the render becomes whatever your prompt says. This node does it on Flux Dev with LoRA support, hosted, at $0.035 per megapixel.

How canny conditioning works

Canny detects hard edges with configurable thresholds and produces thin, high-contrast edge maps - which is why the classic use cases are architecture, mechanical objects and any scene with clear contours. The geometry is extracted from the reference, generation follows the edges, and the prompt decides the materials, the lighting and the style.

The thing people used to ControlNet from the SD days should brace for here: this node has no control weight, no start/end step, no preprocessor choice. There's an image socket, and the endpoint handles the conditioning at its own weighting. A lot of the traditional advice - release the condition after composition forms, weight 0.8–1.2 for strict adherence - is about dials that don't exist on this node. What you control is the input picture, the prompt, and the seed.

Inputs

• image - the edge-condition input. Cleaner source, cleaner guidance.
• prompt - what the shapes should become.
• loras - a JSON string of LoRA entries with weights and scale, merged together. Empty leaves it out. They resolve on fal's side.
• image_size (defaults to square_hd here) plus image_size_width / image_size_height for custom (16–14142, step 16). Set it to match your input's framing unless you want the layout re-cropped around the edges you supplied.
• guidance_scale 3.5 (0–35), num_inference_steps 28 (1–50), seed, num_images 1–4, enable_safety_checker, output_format, acceleration.

The four pack-wide widgets sit at the bottom: max_cost_usd (default 20), reuse_identical_run (on - repeat runs return the paid files), extra_json, max_concurrent.

Outputs

images → preview or save. image_paths points into output/fal/fal-ai_flux-lora-canny/. seed, prompt and info come back for reproducibility and debugging; has_nsfw_concepts tells you whether the safety checker flagged the result, and timings is there if you care how long the wait was.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Manager users: search comfyui-arkennemasis or use Install via Git URL. Restart ComfyUI. No local model downloads, obviously - that's the whole point of the fal nodes.

# .env in the folder with run_nvidia_gpu.bat
FAL_KEY=your-key-from-fal.ai/dashboard/keys

No key box on the node, by design.

Where it goes sideways

Edges ignored. Canny needs contrast. A soft, low-light, hazy reference produces a sparse edge map, and the model is then largely inventing from your prompt. Brighten and sharpen upstream, or pick an input with actual contours in it.

Edges obeyed too literally. Small details in the map can show up as artefacts - the classic "the model faithfully reproduced the flyer stapled to the wall as a weird blob" problem. Simplifying the input (blur, downscale and upscale) helps.

Batching. Treat these sockets as one picture each. If a loader hands you a batch, split it first rather than assuming the endpoint knows what to do with a stack.

Cost per attempt. Iterating on a prompt costs a real 3.5 cents per megapixel each time, and num_images 4 is four times that. Use acceleration while you're searching for the prompt, then turn it back to none for the keepers.