Extract a control map

Reads the structure of a picture and returns a new picture of that structure.

A control map is a picture of the structure of another picture: its outlines, how far away each part is, or where a person's joints are. A request always uses one, and normally the worker makes it for you. There are three reasons to make one on its own:

to check it before you spend kudos on an image, for example that the pose caught both arms;
to use as a picture in its own right: line art to colour in, a depth relief, or flat colour regions to cut a mask from;
to attach it yourself as the for a ControlNet request, so you decide exactly what the new image follows.

Control map: Extracts the selected structure, such as edges, depth, or pose, and returns the map used to guide a ControlNet request.

Guiding a render with ControlNet Source image in the image studio

Examples

The image that was submitted — All the examples are based on this image.

Which one to pick depends on what you want kept. Edges and line art hold the outlines, so a new image redraws the same shapes in another style. Depth and surface direction hold the layout in space and let the drawing go. Pose holds the people and nothing else. Regions give flat fields to work from, either as a plan for a new image or as a mask for another tool. Colour and texture keep the palette or the tones, and the tile detectors keep the picture itself, which is how detail is added or a picture enlarged under guidance.

Edges and outlines

Outlines and the boundaries between light and dark. Canny is thin and literal. The soft-edge detectors follow the shapes a person would draw and ignore fine texture. The scribble detectors draw deliberately rough lines.

cannyFinds the hard edges and draws them as thin white lines on black.
hedDraws softer, hand-drawn-looking edges that follow the shapes rather than every hard boundary.
pidinetDraws soft edges too, keeping fewer of the fine ones than hed does.
teedDraws soft edges with thinner, cleaner strokes than the older soft-edge detectors.
pyracannyFinds hard edges at several scales at once, so large shapes survive alongside small detail.
binaryReduces the picture to pure black and white at a threshold, keeping mass rather than outline.
scribbleReduces the picture to a loose sketch, as though someone had traced it quickly by hand.
fakescribblesTurns detected edges into a scribble, so a photograph comes back looking hand-drawn.
scribble_xdogMakes the scribble from a difference-of-gaussians pass, which keeps more of the fine detail.
scribble_pidinetMakes the scribble from the soft edges pidinet finds, which keeps the shapes and drops the texture.

Line art and straight lines

Clean line drawings, the way an illustrator would ink the picture. mlsd keeps the straight segments only, which is usually the architecture.

standard_lineartDraws the picture as a plain line drawing, black lines on white.
lineartDraws a line drawing with weighted strokes, closer to how an illustrator would ink it.
lineart_animeDraws the line drawing in the flat, even-weight style anime line art uses.
lineart_anime_denoiseThe same anime line drawing with the speckle cleaned out of it.
mlsdFinds the straight segments only, which is the architecture and none of the organic shapes.

Depth

How far each part of the picture is from the camera, drawn light for near and dark for far.

depthReads how far each part of the picture is from the camera, drawn as light near and dark far.
midas_depthReads the same distances with the MiDaS model, which is softer about fine detail.
zoe_depthReads the distances with ZoeDepth, which keeps the relative scale of a scene more consistent.
depth_anythingReads the distances with Depth Anything, which holds edges more crisply than the older models.
depth_anything_v2Reads the distances with the second Depth Anything, the sharpest of the depth readers.

Surface direction

Which way each surface faces, drawn as colour.

normalReads which way each surface faces, drawn as the colours that stand for the three directions.
normal_baeReads the same surface directions with the BAE model, which is steadier on smooth surfaces.

Pose

The people in the picture, drawn as skeletons and nothing else.

openposeFinds the people and draws their skeletons: limbs, joints and the line of the head.

Regions and objects

The picture divided into flat fields, one colour for each kind of thing.

segDivides the picture into regions and colours each one by what kind of thing it is.
oneformer_ade20kDivides the picture into regions using the ADE20K vocabulary, which names indoor and outdoor scenery.
oneformer_cocoDivides the picture into regions using the COCO vocabulary, which names everyday objects and people.

Colour and texture

These keep the palette or the tones of your picture, and let the shapes go. The tile detectors keep the picture itself, as a guide for redrawing it.

colorReduces the picture to blocks of its colours, keeping the palette and losing the drawing.
shuffleShuffles the picture's content around, keeping its colours and textures and losing where they were.
recolor_luminanceReduces the picture to its brightness, read the way the eye weighs the colour channels.
recolor_intensityReduces the picture to its intensity, read as the strongest channel rather than as perceived brightness.
tileKeeps the local detail of the picture as a guide for redrawing it a tile at a time.
tile_ttplanet_guidedThe TTPlanet take on tiling, guided so the redraw follows the original more closely.
tile_ttplanet_simpleThe simpler TTPlanet tiling, which blurs more of the source and leaves the redraw freer.

Open the Alchemy studio