Masks

The Masks tab

The ADJUSTMENTS list at the top of the tab holds Global and then your masks. Click a row to choose it. The sections below edit what you chose.

• The dot before a mask switches it on and off.
• The bin on the chosen row deletes that mask.
• New mask adds an empty mask named "Mask 1", "Mask 2" and so on, and takes the sections back to their top. Its name field has the keys, the name selected, so you can type a name at once; Enter or a click elsewhere ends the naming. A frame can hold up to 16.
• Right-click a mask (Ctrl+click on a Mac) for Duplicate mask, which puts a copy just below it with the same shapes and sliders, named as a new mask is ("Subject 3"), and Copy mask. Once a mask is copied, the menu on any row, Global too, offers Paste mask: see Masks and other frames. Copy masks, next to it, copies every mask at once: Paste settings on other frames then opens with each mask checked and nothing else, so you can uncheck one before pressing Paste.

The list stays put while the sections scroll under it, so Global or another mask is one click away from any slider. It shows up to six rows, fewer in a short window, and scrolls inside itself past them. Click its title to fold it to one line, which names what you are editing.

With a mask chosen, you see, top to bottom:

1. MASK, headed with the mask's name: its name field, Show mask, Invert, and Refine edge, then Shapes, one row per shape in the mask, the chosen shape's settings when it has any, and Add a shape, the buttons that add another shape.
2. The mask's own sliders: WHITE BALANCE, LIGHT, COLOR, CURVES, COLOR MIXER, COLOR GRADING, TINT and DETAIL, each headed with the mask's name as you typed it ("Sky 1 · LIGHT").

Naming, Show mask, Invert, and Refine edge

A mask is named for its first shape ("Subject 1", "Linear 2"). Type in the name field to call it something you'll recognize; a name you type is kept.

Show mask paints the mask red over the picture, stronger where it acts more. It turns on by itself while you are drawing a shape, and goes back to how you had it when you finish.

When you move a slider in the mask's own sections, Show mask turns off by itself, so you see the edit rather than the red. A shape's settings, such as Feather or a range, Invert, Refine edge, and the brush leave it showing, since those are about the shape. Turn it on again whenever you want to check the shape.

Invert flips the whole mask: the adjustment then acts everywhere the shapes are not.

O turns the chosen mask's Show mask on or off.

Rest the pointer on a mask's row in the list and that mask is painted red too, so the one to fix is found among many; on the Retouch tab, a repair's row draws its outline.

Refine edge fits the whole mask's edge to the picture's, after its shapes are combined and inverted. At 0, the default, the mask is exactly what its shapes make. Turned up, where the picture has an edge within reach of the mask's, the mask's edge takes the picture's shape and softness and is drawn toward it; where the picture is flat, the mask's edge is smoothed into a ramp. The slider sets the reach: at 100% about 1.5% of the picture's width each way. It reads the picture's color as well as its lightness, so a red coat against green leaves of the same lightness still holds the mask. It fits an edge drawn near the subject's; it does not find one drawn far from it. It reads the picture from before the Detail section and the capture sharpen, so a Dehaze or Clarity halo is not taken for an edge.

The edge is found on the masks' own grid, 2048 across whatever the picture's size: about 4 pixels a step on a 45 MP frame, so at 1:1 the refined edge is that soft at the least. A Sky made at twice the size is read at that grid too once it is refined.

The shapes list

Each row reads as the way the shape joins (+, − or ∩) and its kind, such as "+ Radial".

• The dot switches the shape on and off without deleting it.
• The half-filled circle inverts that shape alone. It lights when on.
• × deletes the shape. Deleting the last one leaves the mask empty, as New mask makes it.
• Click a row to choose it and see its settings.

Adding shapes

Click New mask, then a button under Add a shape. The first shape always adds. From the second on, the row above the buttons sets how the next shape joins:

Intersect is how you limit one shape to a place, such as a Luminance shape that only reaches the top of the frame: add a Linear, then a Luminance with Intersect.

Press Esc to put a drawing tool down, or click its button again. With no tool in hand, Esc hides a Linear or Radial shape's handles so you can see the picture; click its row or the shape to bring them back.

Linear

Drag across the picture. The adjustment is full where you start and fades to nothing where you let go. Afterwards, drag anywhere in the band between the end lines, or the middle line, to move it, or an end line to turn and stretch it about the middle. Hold Alt while dragging an end line to move that end alone. A drag outside the band moves the picture. Hold Shift during the drag, placing it or moving an end, to keep its angle to steps of 15 degrees. Pressing the middle button during the drag does the same.

Radial

Press at the center and drag outward. Drag anywhere inside the ellipse, or its center handle, to move it; the pointer turns to a move arrow there. Drag a handle at the end of either axis to change that radius and turn the ellipse. Feather (50% to start) is how much of the ellipse fades out toward its edge. Hold Shift during the drag, placing it or moving an axis handle, to keep it a circle. Pressing the middle button during the drag does the same.

Brush

Click Brush and paint. While a brush is chosen, the button reads Paint, and new strokes go into the same brush.

Subtract takes away from the whole mask under the stroke, whatever the shapes before the brush gave there, so a brush over a Sky, Subject, or Luminance shape cuts it. Erase clears only the brush's own strokes, added or subtracted. Choosing Subtract in Add a shape for a new brush puts this row on Subtract until the brush is put down.

Press the middle button during a stroke to make it a straight line from where it began to where you let go.

With the brush in hand, [ and ] make it smaller or bigger, Shift+[ and Shift+] take the feather down or up by 0.05, and Ctrl+[ and Ctrl+] the opacity.

Subject and Background

Subject finds the main subject with a learned model. Background is everything the Subject mask leaves out. Neither needs drawing: one click makes it. A mask with both costs one run of the model.

Sky

Sky finds the sky. It refuses a frame where it finds none rather than inventing one; the status line says "no sky found". With a Sky shape chosen, the button reads Pick: click sky it missed to add it, or right-click what isn't sky to remove it.

It works least well on sky seen through a window or behind dense blossom.

On a frame wider than 6144 pixels, as from a 45 or 100 MP camera, the Sky mask is made at twice the size, so twigs, wires and the rims of leaves against the sky keep a clean edge at 1:1.

Landscape

Landscape, beside Sky, looks over the picture once and offers what it found, largest first, in a menu under the button: Water, Foliage, Mountain, Building and Sky, each with the share of the frame it covers ("Foliage 33%"). Only what it found is offered; with nothing found there is no menu, and the status line says "no landscape found". Choose one to add it to the mask, joined as the row above the buttons says; press Esc to add none.

Each is a shape edged from the picture: water and foliage follow the color line along their edge, mountains and buildings a softer one. Foliage keeps off the sky that Sky finds, so the open sky between branches stays out of it. These shapes take no clicks: to take something out of one, subtract a brush, or intersect it with another shape.

Object

Click Object, then click the thing you want, right-click to exclude something, or drag a box around it. With an Object shape chosen, the button reads Pick and adds more clicks to it.

People

People opens a menu: Whole person at the top, then a person's parts in three groups:

Whole person takes one person head to foot: hair, skin, clothes, and usually what they hold. Where Subject takes everything that stands out, people and things alike, Whole person takes the one person you pick, so on a group you can brighten one person and leave the rest. Where two people stand so close that the model sees them as one, each takes their own side of the pair. Someone of whom only a head shows gets their face. Its edge comes from the Subject model, so loose hair against the sky is kept as Subject keeps it; where the person touches someone else, the line between them is the People model's. Without the Subject model on the machine, Whole person still works with a coarser edge, and the status line says so. The first Whole person you ask for offers the Subject model's download.

Clothing goes by where it is worn: Top takes a coat and the shirt under it as one, Bottoms whatever is worn below the waist. Iris, Lips and Teeth look closer, at each face and then each eye or the mouth, so they take longer than the rest.

On a picture of one person, the shape is made at once, as Subject is, and it is that person's. On a picture of more than one person, each face is outlined and the status line says "click the person, or All people". Click a person to make the shape theirs alone, click All people above the picture to take everyone's, or press Esc to make nothing. On a picture with no one in it, the shape is made at once and is everyone's.

A People shape keeps who it is of, even when it was made on a portrait of one person and you were never asked. On the picture it was made on, that is the person whose face is where it was, however you turn or edit the picture afterwards; a change of white balance, camera profile, lens or noise reduction may ask you to click them once, the face where theirs was outlined more strongly. Pasted or synced onto another picture, it looks for the same person there by their clothes, hair and skin, not by where they stood: pasted onto another frame of the same person, it finds them, though someone dressed very like them can be taken for them.

When it can't be sure, as on most pictures of several people, the mask stays empty and the faces are outlined again, the likeliest person outlined more strongly. When it doesn't find its person at all, as on a portrait of someone else, it never takes whoever is there: the faces are outlined, none more strongly, and the status line says "this mask was made for someone else". Either way, click a person to settle it for that picture, or All people to make that picture's copy everyone's; press Esc, or carry on, and the mask stays empty. If you were on another tab, the outlines come back when you return to Masks. On a picture with no one in it, the mask stays empty and the status line says so. Nothing stops a paste or sync onto many pictures: each one asks when you open it. After a paste or sync onto several pictures, a person clicked on one of them brings up Use this person on all N pictures under the shapes: press it and greycard looks for that person on the other pictures that still have the mask as it came, by their clothes, hair and skin and by where they stood on the picture you clicked, and gives it to them where it finds them (each picture's history gets a step, "Person from" the picture you clicked, that undoes on its own). Where it can't tell, the picture keeps asking, and the status line counts them ("used this person on 6 of 8 pictures; 2 not found"); the cross beside it puts the offer away. An export with a People shape that hasn't found its person writes the mask empty and says so in its result line and the log ("a People mask needs a person picked", or "a People mask's person isn't in this picture").

A People shape made with All people stays everyone's wherever it goes: pasted onto another picture, it takes everyone there.

On a machine without a GPU, a People shape takes several seconds, and an iris, lips or teeth about thirty.

Luminance

A luminance shape takes everything between its Low and High lightness, fading out over Low fade below and High fade above. A new one takes 70 to 100.

The scale runs from 0 (black) to 100, the brightest the camera recorded at a global exposure of 0. Middle grey is about 57. The picture on screen is shown a little brighter than that, so a bright sky can sit in the 70s. A High of 100 takes in everything above it, and a Low of 0 everything below. Dragging Low past High pushes High along, and the other way round.

Color

A color shape takes one range of hues, set by Hue and Width and fading over Hue fade. It takes only colors at least as strong as Chroma, fading out over Chroma fade, so greys stay out.

A new one starts on skin tones (hair of the same hue comes with it), with the dropper in hand: click a color in the picture to center it there, or press Esc to keep skin. Pick gives you the dropper again later.

The dropper reads the mean of a square about the point, drawn around the pointer while you pick. Sample beside Pick sets its side: 5, 11, 21, or 41 pixels of the picture itself, so at fit on a large frame even 41 is a small square on screen, a few screen pixels. Take a wider one on a noisy or textured color. The choice is kept between runs, not in the edit.

What Luminance and Color read

Both read the picture as developed, before any mask's own sliders, the curves or the look, so a mask doesn't shift as you edit under it. While the Detail section does something (texture, clarity, structure, or dehaze), they read the picture from before it and before the capture sharpen, so those sliders don't move them either; with Detail at rest they read the developed picture, sharpen included. Refine edge always finds its edge on the picture from before Detail and before the capture sharpen. Exposure moves them.

A mask's own sliders

Under the mask, LIGHT, COLOR, CURVES, COLOR MIXER, COLOR GRADING, TINT and DETAIL work as they do on the Develop tab, but only where the mask reaches. BLACK & WHITE and the Display curve switch are global only. A mask's tone sliders act even when the global LIGHT section is switched off. Each of the mask's sections has its own switch, TINT included, so you can compare the mask with and without it.

A mask's curve editor shows the picture's histogram behind the curve, as the global one does.

Detail in a mask

The mask's DETAIL has Texture, Clarity, Structure, Dehaze, and Sharpen, the first four as the Develop tab's: smooth the skin of a face and leave the rest sharp, or clear the haze from a distant ridge and leave the foreground. It shows in the viewport, the export, and the thumbnails alike, and its steps land in the history like any other.

Sharpen is a share of the capture sharpen, by the Develop tab's radius, iterations, and threshold: -100 takes the Develop tab's Sharpen off under the mask, +100 puts it on once more; with the Develop tab's Sharpen off, a positive amount sharpens inside the mask alone.

A mask over the whole frame gives the same picture as the global slider, one slider at a time. A mask's Dehaze used together with the same mask's Texture, Clarity, or Structure comes out a little differently from the global pair, since the mask reads the haze before its own detail is applied.

White balance in a mask

Use this for mixed light: a room under lamps with a window, a stage under two colors of light. Set the global white balance for one light, then give a mask the other.

The mask's WHITE BALANCE section is off until you switch it on at its title. The first time, it starts at the global white, so nothing jumps.

• Temperature (2000 to 12000 K) and Tint name the light under the mask.
• Neutral gives you a dropper: click something under the mask that should be grey. It switches the section on for you.
• The switch keeps the mask's light while it is off: off, the area under the mask takes the global white; on again, the mask's own light comes back. Use it to compare the two.
• Double-click Temperature or Tint while the section is on to forget the mask's light. The section switches off, Follow global goes off with it, and the mask follows the global white again; the next time you switch it on it starts there.

There is no As shot or Auto here. With Follow global off, the values are a light: a mask set to 3200 K stays a 3200 K lamp when you move the global white balance. With it on, they are an offset from the global (+25 mired, a tint of +0.002), so a mask made warmer stays that much warmer when the global moves. The offset is in mired, so the same number is the same change of warmth at any global; plus is toward the higher kelvin, as on the Temperature slider. Switching it either way moves nothing on screen, and the dropper still makes the patch grey.

Over blown highlights, a mask whose white is far from the global one can tint them if you pull exposure well down.

Models download on first use

Subject, Background, Sky, Landscape, Object and People use learned models that greycard downloads the first time you need one. A sheet names the model, its size, where it comes from and its license before anything downloads; choose Download or Not now. To get them all at once, use Download all in Settings, Models. The first time a picture opens, greycard asks once whether to fetch every model now; As needed, the default, leaves each to come when it is first used.

Each is downloaded once. The first Subject or Background mask of a session takes a few seconds while the model loads; after that, masks come quickly. Without the Object model, Sky still works, with a coarser edge, and clicks can only remove sky.

A learned mask that a Fill lies over is made again on the retouched picture, so a person a Fill took out leaves the Subject mask too; a Heal, a Clone, or a Fill elsewhere leaves it alone.

On an older processor

The models run on ONNX Runtime, and the copy that comes with greycard needs a processor with AVX2 (most Intel Core processors from 2013 on, and AMD's from 2015 on). On an older one, on Linux or Windows, greycard still starts and edits as usual, but the first time you ask for a learned mask, the eraser or the learned denoiser, the same sheet offers Microsoft's own ONNX Runtime instead: about 9 MB on Linux and 79 MB on Windows, under the MIT license, kept beside the models. It runs the models on the CPU alone, so they are slower than on a GPU, and the status line says so ("on CPU (slower)"). Not now leaves the AI features off for the session, and the status line says why. From the command line, greycard models --fetch onnxruntime fetches it; a greycard export on such a machine before it is fetched writes the picture without the learned masks and exits 2, naming it.

On Windows on ARM

No copy of ONNX Runtime comes with greycard on Windows on ARM, so there it is the same on every processor: the first time you ask for a learned mask, the eraser, or the learned denoiser, the sheet offers Microsoft's ARM64 build, 80 MB from github.com and 16 MB kept, under the MIT license. The models run on the CPU alone, and the status line says so. Not now and greycard models --fetch onnxruntime work as above.

On an Intel Mac

The copy of ONNX Runtime that comes with greycard on an Intel Mac is built without the graphics path the other platforms use, and nothing is downloaded for it. The learned denoiser and the Sky and Landscape masks run through Core ML, on the Mac's GPU or its CPU; Subject, Background, Object, People and LaMa's fill run on the CPU alone. The generative remove needs that graphics path, so its two buttons are greyed, as on a machine without a GPU.

Masks and other frames

Masks are drawn around one picture, so syncing settings onto other frames leaves them out unless you check Masks, or one mask under it by name. Replace masks takes each frame's own away; Add to masks keeps them and puts the copied ones after. See Presets and history.

To take one mask across without the sheet, right-click it in the ADJUSTMENTS list and choose Copy mask, then on another frame right-click any row and choose Paste mask. It goes onto the open frame and every other selected frame, after their own masks, as Add to masks does; a frame that has it already is left as it is. Each frame takes it as one step of its history, "Paste from" the frame it came from. Copy mask leaves the copied settings of Ctrl+C as they were.

A Subject, Background, Sky, Landscape or People mask copied onto another frame is made again for that frame, from that frame's own picture. A paste of masks onto the frame on screen shows at once.