Scenario Detection: The Essentials

Covers model_scenario-detection  |  Provider: Scenario  |  Modality: Image to Image

Scenario Detection extracts structured condition maps from any image. These maps are the input that ControlNet-compatible generation models use to constrain the composition, pose, depth, or structure of a new image. You run Detection first, get the map, then feed that map into a generation model as a guide.

Ten preprocessors are available in a single model. You pick the one that matches what you want to preserve: use Canny to hold the edges, Pose to hold a character's body position, Depth to hold the spatial layout, MLSD to hold the architectural lines. Each detector strips out a different layer of structural information, leaving only what the generation model needs to follow.

The Ten Detectors

	Detector	What it extracts	Best used for
	Canny	Sharp edges and outlines across the full image	Style transfer that preserves layout and composition. Works on any subject.
	Depth	Grayscale depth map: bright foreground, dark background	Recomposing a scene while preserving spatial relationships and perspective.
	Pose	Human skeleton with body, face, and joint keypoints	Transferring or replicating a character's pose. Works across different characters and styles.
	Normal	Surface orientation encoded as color vectors	Preserving fine surface detail and controlling lighting behavior in the output.
	MLSD	Straight line segments only	Architecture, interior design, and technical drawings with strong geometric structure.
	Segmentation	Semantic color regions: each object type gets its own color	Controlling the layout of a scene while changing style, lighting, or content within each region.
	Scribble	Loose, simplified edge lines, similar to a rough hand sketch	Loose style transfer. Also useful as input when you want the generation model to have more creative freedom.
	Sketch	Detailed line drawing preserving more contour information than scribble	Coloring or restyling artwork while keeping the drawn structure.
	Line Art (Anime)	Clean, high-contrast anime-style line extraction	Coloring manga or anime line art, and extracting clean linework from illustrated images.
	Grayscale	High-contrast grayscale with optional background removal	Preparing images for inpainting or masking workflows. Subject isolation.

Which Detector to Use

The right choice depends on how much structural constraint you want and what kind of structure matters for your output.

Preserve the full layout of the image

Use Canny. It captures edges from every element in the scene, including the subject and background, giving the generation model a tight structural template to follow. It works on any image type and is the most universally reliable starting point.

Preserve spatial depth and 3D positioning

Use Depth. The output is a grayscale map where bright areas are close to the camera and dark areas are far away. The generation model uses this to place objects in the right spatial relationship, even if the actual content changes entirely.

Preserve a character's body position

Use Pose. The detector finds the skeleton keypoints of every human figure in the image and outputs them as a simple stick figure map. Feed that map into a generation model and it will produce a character in the same pose, regardless of their appearance, style, or clothing.

Preserve architectural or geometric structure

Use MLSD. It detects only straight lines, which is exactly what you need for building facades, room interiors, floor plans, and any scene dominated by hard edges. It ignores organic shapes entirely, so it will not clutter the map with irrelevant detail.

Preserve the layout of a complex scene with multiple objects

Use Segmentation. It produces a color-coded map where every region is labeled by object type — sky, wall, person, vehicle, plant. The generation model uses this to keep each element in the right place while changing everything else about the visual.

Give the generation model more creative room

Use Scribble. The simplified lines give only a rough outline of shapes, leaving more freedom for the model to interpret the style, texture, and detail. Good when you want the output to feel generative rather than a direct structural copy.

Typical Workflows

Character pose transfer across styles

You have a photograph of a person in a specific pose and want to generate a fantasy illustration character in the same position. Run Detection with Pose to extract the skeleton map. Feed that map into an image generation model alongside a prompt describing the fantasy character. The generated character will match the pose from the original photo.

Architecture restyling

You have a photograph of a building and want to generate an illustration of the same structure in a different architectural style. Run Detection with MLSD to extract the structural lines. Feed the line map into a generation model with a prompt describing the target style. The building's proportions and geometry stay intact while everything else changes.

Scene recomposition with new content

You have an image with a specific spatial layout and want to generate a completely different scene that preserves the depth and positioning of objects. Run Detection with Depth to extract the depth map. Use that map as the ControlNet input for a generation model with a new prompt. The new scene will have objects placed at the same relative distances from camera as the original.

Coloring manga line art

You have a black-and-white manga page and want to generate a colored version. Run Detection with Line Art (Anime) to extract clean linework. Feed the line map into a generation model with a prompt describing the coloring style. The generated image keeps the original linework and adds color within the contours.

Style transfer with full composition lock

You have a reference image and want to generate a new version in a completely different visual style while keeping every element in the same position. Run Detection with Canny to extract all edges. Use the edge map as the ControlNet input with a prompt describing the new style. The composition, framing, and structural layout remain identical.