The depth map

Everything the sheets are cut from is one greyscale picture called a depth map. White is nearest to you, black is farthest, and each flat grey in between is one sheet. Get the depth map right and the rest of the studio is arithmetic; get it wrong and no amount of setting-changing will save the scene.

Two ways to get one

Having trouble coming up with a prompt? Above the prompt box is a shelf of templates, with a Category list to narrow it to one kind — coastal, forest, mountain, mandala, fairy tales, winter, Christmas, Halloween, Easter, Thanksgiving, Day of the Dead, Lunar New Year, the other holidays, weddings, Christian scenes and more. Choose one and it fills in the prompt and the look settings (sheets, shape, fit, detail, the focal subject); a Make it yours field then swaps the one phrase that matters — "lighthouse" for "windmill" — everywhere in the prompt. Each scene template ends by listing what every sheet holds, front to back, one band per sheet, with the subject named in its band. The plan follows that list, so a swapped subject stays on its own sheet. See Plan. You still press Generate Depth Map, and the actual image generated will be uniquely different every time: a card's picture, where it has one, is one example of what that prompt once made, not what you will get. Your material, canvas and finish are never touched by a template. A template goes into the scene that is open. If that scene already has a depth map, the card asks first: Start a new scene with it leaves the open scene exactly as it is, and Replace this scene's prompt puts the template's prompt and settings in place of the scene's — the two are not combined, so the next depth map you generate is the template's scene with nothing of the old one in it. Until you generate, the old depth map stays, and the next build cuts it with the template's settings. To add something to a scene — a moon, a fence — type it into that scene's prompt rather than choosing a template.

Want a whole design rather than a prompt? Finished designs, in the bar at the top, opens one we built and perfected as a scene of your own, depth map and settings included, ready to build.

• Generate Depth Map — one image call, from your prompt. This is the only part of a landscape scene that costs credits, and the button says how many before you press it. The figure follows the Picture shape - a square picture is one price and a tall or wide one another - and there is no quality to choose here: every depth map is made at the quality the site is set to.
• Upload depth map — your own greyscale PNG, JPEG or WebP, white nearest, up to 25 MB and 16.8 megapixels. Free to upload. A picture in colour is read by its brightness; see It came back in colour, and Bringing your own for the size to make it.

What the studio asks the model for

You do not write the brief. From your prompt and your layer count the studio writes a scene plan — a job for every depth, front to back, and the one or two things that must stay recognisable once it is all glued up. That plan becomes the brief: not a picture, but a depth map in exactly that many flat greys.

The centre of that brief is cut-out detail. Anything seen through an opening — a lit window, the daylight between two branches, the gap under an arch — is painted in the grey of whatever is behind it, so a window is simply a piece of a farther sheet showing through a nearer one. That is the whole trick of a papercut lightbox, and it is why a good depth map looks less like a photograph and more like a stencil.

The brief names the detail it wants as things rather than adjectives, and forbids only what spoils the reading of the greys: haze, grain, and above all shading. Image models like to draw a depth map as a photograph of a paper cut-out, with a drop shadow under every layer and a darker rim where two layers meet, and each of those is a gentle slope of grey that the cut has to split somewhere - which is where a ragged edge comes from. So the brief asks for hard-edged flat fills and names every kind of shadow it has seen. Flat fills are also what a coloured illustration is made of, so its first line says that every fill is a neutral grey - equal red, green and blue - and its list of refusals starts with colour, hue, sepia and tint. It never asks for less detail.

Why your greys are not trusted

The picture that comes back is normalised, denoised and cut into exactly the number of depth classes you asked for — but by finding the boundaries between the clusters the model actually painted, rather than by slicing the range evenly. That distinction is worth understanding, because it is the difference between a fence on one sheet and a fence speckled across two.

• A boundary is only ever placed between two clusters, never through one. A fence painted in a grey of its own between the house and the grass is one thing, on one sheet, even when the picture holds more greys than you asked for sheets.
• A shape is read as a whole. Models shade their shapes a little towards the outline, and where that shading crosses a boundary the band along the edge stays with the shape it is inside — rather than eroding the shape on its own sheet and leaving a ring of it on the sheet behind.
• A drop shadow is the surface it falls on. When the model draws a soft shadow round a shape anyway — dark right at the shape's edge, fading back to the plane's own grey a few pixels out — the dark sliver is read as the plane it darkens, not as a farther depth. Read as depth, it used to be cut as a thin slot along the shape's outline: a crack under every step of a lightning bolt, a dashed line along a jaw. A slot drawn with a hard edge on both sides, a hole inside a shape, a dark opening the shadow runs into, and a dark gap between two shapes — the wedge between two petals, the space between two leaves — are all kept as drawn. What tells them apart is how they meet the surface: a shadow fades into it gently, over several pixels, and the edge of a gap, however soft, steps all the way in two or three.
• A plain page round a design is the backdrop, not the nearest depth. White is nearest, so a design drawn on white paper would otherwise stand on a page read as the front sheet — a board with the design set into it. One flat grey running right round the picture's edge and in to the design is read as the back sheet instead, and the design is cut at the depths in front of it — see A page round the design, which also says how to keep the page.
• A plane that drifts across itself stays one plane. A roof painted a little lighter at the ridge than at the eaves, with no edge anywhere on it, is one roof on one sheet, not a house with the top of its roof on the sheet in front.

Only a real step in the picture, or a boundary that reaches well past the cut on both sides, divides a region. Each class that survives is one candidate sheet.

Bringing your own

Upload depth map takes a PNG, a JPEG or a WebP of up to 25 MB, greyscale, with white nearest. Flat, distinct greys work best — a hand-painted map with six deliberate tones will beat anything automatic. Whatever format it arrives in, it is kept as the greys the build reads — a greyscale PNG — so a grey map keeps every grey it had, and one in colour is kept as its greys, the same as a generated one; see It came back in colour. Anything transparent is read as black, the farthest depth, so fill a transparent background with the grey you mean before you upload.

A portrait, or any tone map where the depths are shades of a face rather than planes of a scene, builds best with few layers - three or four - and with the subject named as the Focal subject, so the studio protects it. Every soft shade becomes a sheet edge, so six layers of a face is hundreds of loose pieces; the sheets are kept as drawn rather than flipped into boards, and those pieces are cut free for you to glue. See Every sheet is read both ways round.

An uploaded map is charged as a picture on pay as you go, once, the first time it is built — the same rule as a generated one. Rebuilding it costs nothing.

Its shape comes from the file, not from Picture

An uploaded depth map does not have to match the Picture setting - portrait, landscape or square. A wide file makes a wide picture and a square file a square one, and the border is then applied to it as to any other. Picture neither crops nor stretches an upload: it is the shape of a picture the studio generates. Changing it, the prompt or Layers after an upload still raises the depth map made for other settings · regenerate badge, and for an upload the badge can be ignored, since nothing about the picture was asked of the model.

What size to make it, in pixels

About the size the studio generates its own depth maps at, or larger: 1536 × 1024 pixels for a landscape (wide) picture, 1024 × 1536 for a portrait (tall) one, 1024 × 1024 for a square. The build works on the file's own pixels without resizing them, so the pixel count decides how fine a line survives: a line under about three pixels wide may drop out, which is about half a millimetre on a 300 mm scene cut from a 1536-pixel picture. A larger file keeps finer lines and takes longer to build: a scene at twice the width and height, four times the pixels, took nearly three times as long.

The largest it takes is 16.8 megapixels - 4096 × 4096, or as many pixels in another shape; a file with more is refused with its size, before anything is built from it, and needs scaling down. The same limit holds for the depth map in a .zip opened with Open scene, when the scene takes its picture from the file. A scene you already have with a larger depth map is not affected: it builds as before.

Getting the picture you wanted

It came back in colour

Now and then the image model answers a brief for greys with a coloured illustration: a papercut in cream, orange and tan on black, say, where each ring should have been a grey of its own. The build only ever reads a picture's brightness, so the studio keeps the picture as the greys it will actually be cut from and shows you those, with a note above it: The image model drew this in colour, not in greys (for a picture you uploaded: This picture arrived in colour). What is on the canvas, the depth.png in the export and what the sheets are cut from are then one and the same picture.

What colour costs is separation. Two colours of about the same brightness become two greys only a few levels apart — in one wreath of flowers, the orange blooms and the tan heart came out eight levels apart out of 255 — and the build puts them on the same sheet, or divides them in an order the prompt never asked for. So look at the greys before you build:

• If every part you wanted at its own depth has a grey of its own, build it. Nothing is lost: the colours were never going to be read.
• If parts you wanted apart are the same grey, generate again — it is charged like any other generation. The brief asks for neutral grey in its first line and again in its refusals, and most pictures come back grey. For an upload, repaint it in greys and upload it again, which is free.
• If the build is refused with "holds 4 distinguishable depths, not the 5 asked for" and the note is above the picture, the colours are probably why. Setting Layers to that number (the message calls it Sheets) will build, but check the order in the stacked view: the sheets follow the colours' brightness, not the depths your prompt described.

The note stays with the picture — after a reload, and in Previous depth maps when you bring it back — and goes when another depth map replaces it. A picture counts as colour when a patch of it bigger than about one pixel in two thousand has its red, green and blue 20 or more levels apart, which no grey map comes near, compressed or not; a pale sage beside a dusty rose is enough. A scene whose picture came back in colour before the studio looked for it is converted, and gets the note, the first time it is opened — nothing is rebuilt or charged for that, because the greys are the ones it was always cut from.

It came back plain

Raise Detail. Simple deliberately asks the model to hold back; Medium gives every depth a few clear openings; Detailed asks for the whole papercut lightbox — rows of windows and door frames, railings, branch structure with daylight between the branches, small figures. A Detailed picture that still comes back plain is one to generate again, with the things you want named in the prompt. Models are much better at drawing a thing you named than at inventing one to fill a gap.

It is cut off at the edges

The design ran off the sides of the picture. That is Fit on Edge to edge, which asks the model for a scene that reaches every edge, and a mandala or any single object is not a scene. Set it to Whole design, nothing cut off and generate again: the brief then asks for the design complete inside the picture with a margin. Once it is set, Around the design appears beneath it and decides whether the space round a round design stays plain, fills with matching ornament, or gets a patterned border — the Outline border cuts that space away again. The mandala and floral templates set the fit for you, with the Outline border.

There is a halo or a ring around something

That is shading around an outline being read as its own depth. Shading inside a shape's outline stays with the shape, and a drop shadow outside it is read as the surface it falls on — see Why your greys are not trusted — so thin cracks along the edges of shapes should not appear. A shadow wider than about 3 mm, or one that meets the surface more steeply than a gentle fade, is read as drawn, so a heavily shaded picture can still produce a ring. Generating again is quicker than fighting it.

The subject is wrong, or from the last scene

Check Focal subject. It stays with the scene, and the brief asks the model for it whether or not your prompt still mentions it — so a scene whose prompt changed from a deer to a castle can still be asked for a deer. The panel says when the focal subject names something the prompt does not. A new scene starts with it empty, which is the safest place to leave it unless you have a reason.

Previous depth maps

Whatever replaces the depth map on the canvas — an upload or a new generation — the one it replaced is kept. The last eight sit under the buttons as Previous depth maps, and clicking one brings it back, keeping the current one in its place. Nothing is lost by trying another generation, and the line under Generate Depth Map says so whenever the scene already has a depth map.

Starting something new

The prompt box belongs to the scene that is open, and what you type is saved into it. There is no "save as": a new idea typed over an old scene rewrites that scene, and the next generation replaces its picture. To start something different, press New scene beside the scene's name at the top first. Once the prompt no longer matches the one the open scene's depth map was made from, a note under the prompt box says so and offers New scene with this prompt. That takes the prompt and settings on screen to a new scene and puts the open one back to its old prompt, exactly as it was with its depth map. Changing the prompt of the open scene on purpose — a bigger moon, one more tree — needs nothing: carry on and generate.

A map brought back needs a rebuild, because the sheets on screen belong to the map before it. The studio says so rather than showing you sheets that do not match the picture.

Which way round the picture is read

White is nearest: that is what the brief asks for, and how every picture is read unless one of two things turns it round.

• A scene drawn edge to edge gets nearer toward the bottom of the frame. If the bottom of the picture comes back noticeably darker than the top, the studio concludes the model drew it dark-near and inverts the whole thing so white is nearest again. The Debug tab says so: its badge reads orientation flipped, and the note under it gives the two brightness figures that decided it. An uploaded landscape painted black-near therefore needs no inverting by hand.
• A design drawn whole — with Fit on Whole design, nothing cut off — is an object, not a landscape, and its top and bottom say nothing about which way round it was painted: a heart is wide at the top and pointed at the bottom whichever way it was drawn. It is always read as the brief asks, white nearest.

When a build comes out back to front anyway — what should be the front sheet is at the back — turn on Reverse the depth order under the border and frame settings and build again. The picture is then read the other way round, black nearest, and the Debug tab's note says so. It is the way round for a map you painted or exported black-near, and for the occasional scene whose nearest things are at the top of the frame, such as branches overhanging the view. It never reaches the brief, so it is a free rebuild of the same picture. Inverting the file before uploading it does not do the same thing, so use the switch.

When a depth map is refused

Three refusals are worth recognising, because all three are about the picture rather than about the studio.

• It could not be decoded as an image. Supply a PNG, a JPEG or a WebP.
• It holds fewer distinct greys than you asked for sheets. A four-tone picture cannot be cut into eight depths. Ask for fewer layers, or paint more tones.
• One depth class holds almost none of the picture. The model drew fewer distinguishable depths than you asked for sheets. The refusal says how many it did draw - "holds 5 distinguishable depths, not the 6 asked for" - and the fix is under it: set Layers to that number - the message calls it Sheets - and press Build Scene again. The depth map is kept, so that rebuild costs nothing. Generate again if you want the full count.

All three are much more common with an uploaded map than a generated one, and all three are cheap to recover from: nothing is charged for a refusal. A generated map refused for its count can be one that was drawn in colour: if the note is above the picture, see It came back in colour.

When the picture no longer matches the settings

The bottom bar raises depth map made for other settings · regenerate when you change something that went into the brief — the prompt, the layer count, the detail level, the Picture (a square, a tall and a wide picture are different pictures), the focal subject, the fit and what goes around a whole design — after the picture was made. The border is not one of them: it is chosen after the picture exists and never reaches the brief, so changing it is a free rebuild and raises nothing. The scene keeps the settings its picture was made with, so the badge is still there after a reload or when you open the scene again. It is advice rather than a refusal. Building anyway re-cuts the existing picture under the new settings, which is free and often fine; regenerating composes a new picture for them, which costs an image call.