SD 1.5 ControlNet Canny (lllyasviel): edge-map input, OpenCV 100/200 thresholds, UniPC 20 steps

ControlNet ("Adding Conditional Control to Text-to-Image Diffusion Models", Lvmin Zhang and Maneesh Agrawala, arXiv 2302.05543) adds an extra conditioning input to a pretrained diffusion model. lllyasviel/sd-controlnet-canny is the Stable Diffusion 1.5 checkpoint conditioned on Canny edges; it was trained on 3M edge-image/caption pairs for 600 GPU-hours on A100 80 GB with SD 1.5 as the base. Use it with SD 1.5; it also works experimentally with other SD 1.5 models such as DreamBooth fine-tunes. Control image: a monochrome image with white edges on a black background. The card's example builds it with OpenCV (pip install opencv-contrib-python): cv2.Canny with low threshold 100 and high threshold 200, then stacks the single channel into three. diffusers recipe from the card: ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", fp16), StableDiffusionControlNetPipeline from runwayml/stable-diffusion-v1-5 with that controlnet (safety_checker None in the example), UniPCMultistepScheduler, optional enable_xformers_memory_efficient_attention() and enable_model_cpu_offload(), then pipe(prompt, image, num_inference_steps=20). The same authors released sibling SD 1.5 ControlNets for depth (MiDaS), HED soft edges, M-LSD straight lines, normal maps, OpenPose, scribbles and ADE20K segmentation. Licence: CreativeML OpenRAIL-M. For SDXL there are separate Canny ControlNets (for example xinsir's). Based on the lllyasviel/sd-controlnet-canny card pinned to revision 7f2f69197050967007f6bbd23ab5e52f0384162a and observed on 2026-10-28. The card gives no conditioning-scale guidance; nothing was run. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/lllyasviel/sd-controlnet-canny/raw/7f2f69197050967007f6bbd23ab5e52f0384162a/README.md",
      "note": "Retained source snapshot observed 2026-10-28; TEXT.txt; UTF-8 bytes 391:468, 9342:9618, 9509:9618, 10282:10322, 10302:10322, 10774:10849, 11111:11172, 11349:11519, 4:259; exact excerpt SHA256 4bd16504124a6ab9bcd112929d780c073b2503e5131cd4f17419cdb34733c54e,3d4078871e1ab0228b689b2191be14408cee8e1e2d59d42659cc2364ca5c14c9,89f3ef798d7d48aa2a94ef77"
    }
  ]
}