Docs / The pipeline / Locate & Segment

Locate & Segment

Step 3 · Locate & Segment is where a flat scene becomes a stack of individually masked elements — one clean cut-out per object, each ready to become a sprite. You work in two AI passes: find objects reads the picture and lists what is in it as tags, then segment traces a pixel-perfect mask around every instance of those tags. Everything the AI finds — plus anything you compose or hand-edit — collects in one shared Found Elements tree over a live Background remainder. You are done when every object you care about is its own element and the Background holds only what stays behind them; Confirm & Proceed carries the stack to Step 4 · Background.

Costs. Two actions spend credits, and both print the price inside the button before you click. Find Objects costs a credit per run. Segment costs one credit per tag it is masking — segmenting four tags costs four — with a minimum of one. Everything else is free: drawing, resizing, renaming and deleting regions; highlighting tags; grouping and nesting; hiding; keeping an element in the background; hand-editing a mask; and composing a new element in Build. A confirm dialog shows the cost once more before any credits leave your balance. See Credits & plans.

The workspace and its boxes

The left panel, Detected Bounding Boxes, shows your scene with a colored box around every detected element. The help line under the heading covers navigation: scroll to zoom at the cursor, middle-drag or Alt+drag to pan, and click to cycle overlapping boxes — smallest first, one more click to deselect. Top-right, the {NN}% button prints the zoom and resets it to 100%, and a {n} Sprites Found chip tracks the total.

A box is colored by its tag and labeled with its name — a solid outline when masked, dashed while still a prompt. Hover one to float its cut-out with a green glow and to reveal a red × that deletes the box and its element (the pixels return to the background). Double-click a box to rename it inline; Ctrl-click anywhere on the art toggles an amber multi-selection for grouping; clicking the empty letterbox clears everything. Once you hide an element (below), swatches appear up here — Transparent (checker), Dark, Light, Blue, plus show all — to set what shows through the holes.

The Step 3 workspace after auto-detection: colored boxes over the scene on the left, the Scene tags and Found Elements list on the right.

Note. Once an element is masked, its box belongs to the segmentation, so resize handles appear only on the rare unmasked prompt box (mostly older projects). Dragging a fresh rectangle now creates a Manual Region, not a bare box.

Scene: tag the whole picture

The Scene card drives detection across the whole image. Work top to bottom.

Find Objects — the first AI pass. It lists what it sees as tags; the button then reads Find More Objects, appending new tags without disturbing the ones you have. The gear opens Find Objects settings: an AI model picker (each option prints its cost, one marked recommended) and a Hide already-found elements from the analysis toggle that cuts objects you already found out of the image the AI studies, so it concentrates on what is left.
Curate the chips — a filled chip has masks; a hollow, dashed chip is not segmented yet. Click a filled chip to highlight its masks (Ctrl-click for several); its trailing × removes the tag and every element it found. The dashed + chip lets you add your own tag for anything missed — typed tags segment just like found ones.
Segment — the second pass, which traces the masks. Its label names the job and the cost: Segment 4 tags (4 cr), or Find more "barrel" (1 cr) when one tag is selected and you want more of it. Selection beats hollow tags, which beat all tags.

The card's SOLO button narrows Found Elements to only the Scene's results.

The Scene card: tag the picture, curate the chips, then segment — hollow chips are not masked yet.

Tip. To hunt for more of one thing — every barrel in a stack — select that filled tag before pressing Segment. With nothing selected, Segment masks the hollow (new) tags first.

Careful. A run segments at most 24 tags; past that the card warns Only the first 24 tags are segmented … and grays the over-cap chips. A fruitless Find More Objects reports No additional objects found …, and a tag that found nothing becomes a struck-through nothing new found chip that later runs skip until you click it to re-enable.

Segmentation settings

The gear next to Segment opens settings that apply to every run, Scene or region — only the re-segment button is scoped to its own card.

Min confidence — the score a mask must clear to be kept (default 0.35). Lower it to catch faint objects, raise it to reject junk.
Max instances per tag — the cap on how many copies of one tag the AI returns (default 32, which is also the most you can set). Lower it to stop a tag from over-splitting into many pieces.
Input upscale (sharper edges) — 1×, 2×, or 4×. Upscaling the input sharpens mask edges, which pays off on small or pixel-art elements.
↻ Re-segment all from scratch — wipes the auto-found elements of the current tags and masks the image again at the same price. Merged and hand-edited elements survive, so only automatic cuts are re-rolled.

Segmentation settings — confidence, instances per tag, and input upscale are shared across every run.

Careful. A 4× upscale on a large scene can exceed the size limit and return Image too large for segmentation — lower the upscale factor. — drop to 2× and it goes through.

Manual Regions: a focused search

When one corner of the scene is busy, draw a region and search just that crop. Each region is a mini-Scene.

Add Region — the button flips to Drag on image…; drag a rectangle and it becomes Region 1, with its own Find Objects, tags, Segment, and gears.
Aim and adjust — click a region's outline on the canvas to select it, then drag its corner circles to resize. In the card, the header highlights its elements, the chevron collapses it, the pencil renames it, SOLO isolates its elements, and the trash deletes the region and everything it found.
Segment the crop — the crop is upscaled to fill the frame, so the AI sees more detail than at full-scene scale and the masks come back sharper.

Each region's settings add a Segmenter menu the Scene lacks. Its default, finds your tags, works exactly like the Scene. The alternative, everything in the area, carves the crop into every object it can see and ignores tags — its button reads Segment everything here (1 cr) for a flat credit, ideal for UI panels or dense clusters where naming each object would be tedious. That mode hides both Min confidence and Max instances per tag — confidence filtering does not apply here — and offers a Detail (finer finds smaller pieces) choice instead: coarse, normal, or fine. Only Detail and Input upscale (sharper edges) remain.

A manual region drawn over part of the scene, with its own tag-and-segment card for a sharper, focused search.

Note. The everything in the area segmenter is region-only — with no sense of names it would be noise across a whole scene. A region under about 16 pixels refuses to run: This region is too small to analyze — draw it a bit larger.

Build: compose an element for free

Build in the workspace header (it turns teal) composes a new element from masks you already have — no AI, no credits. While it is on, boxes stop catching clicks so every click lands pixel-accurately, and a composer bar docks under the canvas.

Left-click a mask to union its pixels in — they tint emerald.
Right-click (or Ctrl-click) a mask to subtract it — those pixels tint rose, and subtraction always wins.
Click empty space for the background remainder there: the whole scene minus every mask, or just a region's leftover when you click inside one.
Name it and Create element — the docked bar shows a live cropped preview and +adds −subs counts. Name the result or accept Element N; your source masks stay untouched.

The × clears your current picks, or exits Build when there are none — Esc does the same.

Build mode: click masks to union them (emerald), right-click to subtract, and compose a new element for free in the docked bar.

Tip. Build is the free way to fix segmentation that came back over-split or fused — merge two halves of one object, or carve one out of a shared blob, without spending a credit. A composed element is protected, so Step 4's keyer and any later re-segmentation leave it alone.

Found Elements, groups, and the Background

Everything the Scene and your regions find lands in one shared, collapsible Found Elements tree. Hover a row to preview its mask on the canvas. The count reads (n), or (shown/total) while a solo is active.

To group, Ctrl-click a set of elements — in the list or on the canvas — then press Group (its label counts your selection) to fold them into a Photoshop-style folder. Groups nest, and dragging a row onto a group drops it inside. Every row carries the same actions, at any depth:

Icon	What it does
pencil	Rename the element or group inline.
ungroup	Groups only — dissolve the folder, moving its members up one level.
lift-out	Children only — take one element out of its group.
mountain	`Keep in background` — the element's pixels stay in the background image while it stays a normal layer, so it renders both behind and in front. Green when on; it is not a folder.
eye	Hide the element to peek at the background underneath.
brush	`Edit mask` — hand-add or -remove pixels; the edit is protected from later re-keying.
trash	Delete the element, or a whole group subtree — pixels return to the background, no undo. Use ungroup to keep the members and drop only the folder.

At the bottom, the dashed Background row is the live remainder — your scene minus every cut-out, with kept-in-background elements left in. Its subtitle counts them (scene − 5 elements · 1 kept in); hover to preview it, and it exports as your PSD's bottom layer. A group has no mask of its own — masks live on the leaf elements.

Found Elements collects every result into a nesting layer tree; the dashed Background row is the live remainder that exports as the bottom layer.

Confirm & Proceed

When every object is its own element and the Background holds only what belongs behind them, click Confirm & Proceed to move to Step 4 · Background, where those holes are repainted into a seamless backdrop. The button is disabled only while the Scene's Find Objects run is in flight — a region's own tagging run leaves it enabled. You can always come back — segmenting more objects later just punches fresh holes for Step 4 to fill.