Image Control - ControlNet XL Canny

What it does

ControlNet XL Canny generates a new image that matches the outlines of a canny edge map while following a text prompt for the style, mood, and content. It is a ControlNet adapter on top of Stable Diffusion XL (SDXL): instead of leaving SDXL to imagine the layout from scratch, you give it the edge structure of a reference photo (or a sketch / line art), and SDXL fills in the rest. The result is image generation with strong structural control, same composition, completely new style. This is a dedicated canny model (trained only for canny), which gives sharper structural adherence than the more general ControlNet-Union variant on edge-based tasks. The xinsir release was trained on 10M+ carefully captioned images and outperforms other open-source canny models (diffusers/controlnet-canny-sdxl-1.0, TheMistoAI/MistoLine) on both aesthetic score and perceptual control.

Problem it solves

• Composition-locked generation – Keep the exact layout, perspective, and outlines of a reference image while changing everything else (style, lighting, materials, era)
• Sketch / line art to image – Turn a hand-drawn sketch or a clean line drawing into a fully rendered image
• Style transfer with structure intact – Reimagine a scene in watercolor, anime, photorealism, etc., without losing the original geometry
• Architectural / product visualization – Generate alternate renderings of a building, room, or product where the structural lines must be respected
• Stronger control than Union-Canny – When the structural fidelity of edges matters most, this dedicated model adheres more sharply than the general Union variant
• High-resolution output – Native 1024×1024 generation that matches SDXL quality

Input/Output

• Input:
• Canny edge map (grayscale image with white edges on black background) — typically produced by the Canny Edge Annotator node from a source photo
• Text prompt describing the desired image (style, content, mood)
• Negative prompt (optional) — what to avoid (e.g., "blurry, low quality, watermark")
• Options: Strength of canny edge guidance (0.0 – 2.0, default 1.0)
• 0.0 = ignore the edges entirely · 2.0 = strictly follow them
• 0.8 – 1.2 works well for most structural tasks
• Number of diffusion steps (20 – 30, default 25)
• 20 – 30 gives good quality
• Classifier-free guidance (CFG) scale (default 5)
• 5.0 – 7.5 is typical for SDXL
• Random seed (-1 for random, or a fixed integer for reproducibility)
• Output image width (512 / 768 / 1024) — the control image is resized to match
• Output image height (512 / 768 / 1024) — the control image is resized to match
• Advanced — Negative prompt (leave empty for no negative guidance)
• Output: Generated image
• Up to 1024 × 1024 (native SDXL resolution; bucket resolutions also supported)
• Photorealistic or stylized depending on the prompt
• Structure follows the canny edge map; everything else follows the text prompt

Accuracy & Speed

Technical Details

Compliance & Provenance

For more on how we classify models and mark outputs, see our AI Policy.

Model Source

• HuggingFace: https://huggingface.co/xinsir/controlnet-canny-sdxl-1.0
• License: apache-2.0