Manual de Stacker

Align

Align warps every image onto the reference so that the only thing that changes through the stack is focus. Without it, the small magnification change that comes with every focus step, the drift of a rail, or the wander of a hand-held camera would put the same detail at different places in different frames, and the stack would smear every edge.

The step is recommended, not required. If you shot on a rigid rig and the images are already in register, Stack will run without it; but even a tripod bracket changes magnification a little from frame to frame, so for almost every stack the right order is Align, then Crop, then Stack.

Alignment is one job with a handful of choices in front of it, and the panel is laid out in that order: the choices live in the Method, Model and Options sections at the top and are set once per way of shooting; most people pick a Method and never open the two sections under it. The run itself is one button in the Alignment section beneath them, and Diagnostics reports what every image needed afterwards.

Method

One preset per way of shooting. Picking one sets the Model and the Options below to the combination that suits that capture, and the sentence under the row says what the chosen preset does. Change any of them by hand and the preset reads Custom; set them back and it reads the preset again. The choice is remembered as your default for the next stack, and it applies on the next Align images run (the local field, which runs at crop time, on the next Crop all images).

In terms of the five switches under Options, every preset starts from the same base: the Similarity model, Drift correction against the reference on, Joint-sharpness weights on, and the other three off. Then Rail and Live subject turn on Local residual field at Crop, Handheld turns on Rotation / scale search (handheld), and Microscope turns on Blur matching. Tripod is the base itself, and the default for a new installation.

Which to pick:

• Tripod for almost everything shot on a rig: the camera’s own focus bracketing, a focus ring, a focus rail with a fixed camera. Start here even if you are not sure.
• Rail when the camera itself travels toward the subject, so that near parts of the subject grow faster than far parts between frames. A single global warp cannot fit that; the local field at Crop can.
• Handheld when the frames are visibly shifted or turned against each other. It costs a little time and loosens the plausibility checks, so use it only when needed.
• Live subject for a bracket where something moved: a leg, a stem in a breeze, a specimen that settled. The local field follows the movement only where both images are sharp, so the soft background is not dragged with it.
• Microscope for stacks where the first and last frames are so far apart in focus that each is a blur in the other. Blur matching removes the bias that mismatch causes, at some cost in time.

Rail and Live subject set the same switches. The program keeps whichever you picked, so the row still says what you meant.

Tip. If a run with Tripod reports fallbacks or outliers on a hand-held bracket, switch to Handheld and run again; if it reports them on a rig, look at the Diagnostics table first. A single image out of order shows there and is fixed by Reorder by scale, not by a different preset.

Compare methods

Not sure which preset fits the capture? Compare methods…, under the presets, aligns the stack every way at once and lets you compare the results in the viewer, the way the adjustment matrix tries every engine. Each way of aligning that solves differently is run in full, exactly as Align images would run it: Tripod (which also stands for Rail and Live subject, whose local field is applied at crop time and so does not change the solve), Handheld, Microscope, Tripod with the Affine model, and Tripod without the drift correction. A Custom combination of your own that matches none of them is tried too, first. Each result is then stacked small with the default engine, so you can compare pictures rather than numbers.

While the run goes on, the stage is the run’s: the viewer shows Aligning pixels… with a bar for its progress across every method and how many methods are done. The Images panel and the bar above the viewer stay away until every method has aligned.

Then the Align stage becomes the comparison. The Images panel is titled Alignment methods and lists the methods alone, one row per method in the same form as the image rows, or one card per method in the strip under the picture. The unaligned images wait, since the comparison has nothing to do with them. Each method shows its small picture, how many images it adjusted and its mean residual, and its symbols: a green check for a clean solve, the fallback symbol with a count when images fell back, and the arrows when the scale trend is not monotone. The one-line explanation is on hover. A gold check marks the Method chosen now. The arrow keys step through the methods.

The viewer shows the methods large, each stacked at up to 2048 pixels on its long edge, enough to judge edges at 100 %. The method shown is A, marked gold where an image’s checkbox sits. Shift-click another method to compare against it: that is B, marked blue. Shift with the arrow keys steps B. With no B chosen, the method after A stands in. A bar above the viewer holds the rest:

• Single shows A alone. Hold to flip, or holding X, shows B in its place at the same zoom, so whatever changes is the alignment.
• Side by side shows A and B in two panes that zoom and pan together.
• Wipe shows one picture split by a divider, A on the left and B on the right. Drag the divider to sweep between them; drag anywhere else to move the picture.
• A · B: A on the gold half, B on the blue half. The knob between them swaps the two, its arrows turning once round.
• Fit and 100% set the zoom, which stays put as you step through the methods.
• Stop, on the tracker’s Align card, ends a run early and keeps the methods that have landed.
• Use adopts the method shown: its switches become your Method, its alignment lands as if you had pressed Align images, and Crop is next.

Use ends the comparison and the Images panel goes back to the unaligned images. So does Esc, without changing the Method. The results stay until the stack changes: adding, excluding or reordering an image, or choosing a new reference, voids them. Until then Show results under the presets brings the comparison back as you left it.

The button is available whenever Align images is: every image developed, two or more of them, nothing cropped yet, nothing running. A run takes about one alignment per method, and the tracker’s card counts the images through the whole run.

Model

The motion the solver fits for each image, as a drop-down.

Why Similarity is the default. Shear has nothing physical to fit in a focus bracket. Given the freedom, the two extra terms absorb the difference in blur between the images and any parallax instead, and the fit gets worse, not better. The alignment audit measured this: on a test stack the similarity fit’s worst-image error was well under the affine fit’s, with nothing else changed. Choose Affine only when you know the image really is stretched.

Changing the model by hand sets the preset to Custom.

Options

The five switches behind the presets, each one row with its explanation on hover and each a stop of the Align tour. You do not need them unless a preset is not quite right for a stack. Changing one sets the preset to Custom unless the result is exactly another preset.

It helps to know how the solver works, in two sentences. Each image is first aligned to its neighbour in the stack and those small steps are added up toward the reference; this chain is the main estimate, because neighbouring images have nearly the same focus and nearly the same blur. Then, optionally, each image is also solved directly against the reference, and that answer is used only if it stays close to the chain.

Why the chain comes first. The images in a focus stack differ in blur, not only in position. A solver that compares pixels will explain a blur mismatch as motion: a blurred edge reads as a slightly larger one, so the size of a blurrier image is over-estimated. The images farthest from the reference in focus are the ones most blurred against it, and would be the worst solved. Aligning neighbours keeps the blur mismatch small on every step.

Drift correction against the reference

“After chaining adjacent images, re-solve each image against the reference; accepted only within the drift caps. Off = pure chain.”

On, the chain’s small errors cannot add up through a long stack, because every image is checked against the reference and nudged if the two answers disagree slightly. The nudge is accepted only when it is small: for the rig presets, within one per cent of scale, half a degree of rotation and two per cent of the frame in shift; for Handheld, roughly twice as much. A direct answer outside those caps is thrown away and the image keeps the chained estimate, which the run reports as a fallback. Off, the chain is the answer, which is how some other stacking programs work; use it if the direct solve keeps falling back on a stack you trust.

On in every preset.

Rotation / scale search (handheld)

“Log-polar Fourier–Mellin seed before phase correlation, and wide plausibility caps. Off for rigs — it can only add noise there.”

The fine solve has to start close to the right answer. On a rig it starts from “no motion” and a coarse shift search, which is enough. On a hand-held bracket the images may be turned by many degrees and differ in size, so this switch adds a coarse search for rotation and scale before that, and widens every plausibility check (up to twenty degrees of rotation, and about a third of the frame in shift, instead of three degrees and fifteen per cent). On a rig the search has nothing to find and can only land on a wrong answer, so leave it off there.

On in Handheld only.

Joint-sharpness weights

“Pixels sharp in one image and defocused in the other stop voting — the blur-mismatch bias measured in the audit.”

Every pixel contributes to the fit in proportion to how similar its sharpness is in the two images being compared. A pixel that is sharp in one and defocused in the other carries the blur-mismatch bias explained above, so it is given little say. Leave it on; the cost is small and the gain was measured.

On in every preset.

Local residual field at Crop

“After the global fit, a stiff per-node displacement grid (48 px, ≤3 px) restricted to regions sharp in both images. Rail parallax, specimens that settled. CPU warp.”

A global warp moves the whole image as one. This switch adds a second, small correction on top: a grid with one node every 48 pixels, each allowed to move the image under it by at most 3 pixels, solved only where both images are sharp so that soft areas are left alone. It is applied when the crop is baked, not during the Align run, so turning it on after a run takes effect at the next Crop all images. It is the fix for parallax on a rail and for a subject that shifted slightly between frames. The warp runs on the processor rather than the graphics card, so the crop takes longer.

On in Rail and Live subject.

Blur matching

“Estimate the relative defocus per image and blur the sharper side to match before the direct solve. One extra Gaussian per image.”

Before the direct solve against the reference, the program estimates how much blurrier one image is than the other and blurs the sharper one by that amount, so the solver compares like with like. This removes the blur-mismatch bias at its root instead of weighting around it, and in the audit it gave the smallest worst-image error of any setting, but it adds a blur pass per image and makes the run noticeably slower. Worth it when the far ends of the stack are very defocused against the reference, which is the microscope case.

On in Microscope only.

Advanced

One slider, Align seed resolution (px), with a stepper: “Phase-correlation seed resolution; raise for very high-res sensors.” It is the size of the coarse first pass that finds the rough shift before the fine solve starts, from 192 to 1536 pixels in steps of 32. The default is 384. Raise it if a very high-resolution sensor’s images come out with a fallback or a wrong coarse shift; otherwise leave it. The note under it says it applies on the next Align run. The section’s caption lights up when the value is not the default.

Alignment

Align images

Align images runs the alignment on every image in the stack, using the Method, Model and Options above. The button is available once every image has developed and there are at least two of them.

While it runs, the step card at the top shows the progress as a count of images and a time left. Pause holds at the next image, and Resume carries on. Stop ends the run at the next image; an alignment is one answer for the whole stack, so a stopped run leaves the images as they were before you pressed the button.

When the run lands, four things happen at once:

• the status bar reads “Alignment complete”;
• Show aligned images turns on, so the viewer shows the registered images;
• the crop box on the next step is set to the automatic rectangle;
• Diagnostics unfolds to show the run.

Any stack results you had are dropped, because the pixels every engine would see have changed. Press Align images again whenever you change the Method or an Option, or change the reference image or the order: a changed setup applies on the next run, not to the one already in.

If you come back to Align after a crop has been baked, the button means something more: the images on disk are already warped and cropped, so aligning them again would warp them twice. Align images first puts the full-size images back, then aligns those, and the status bar says “Undoing the crop, then aligning”. That needs the uncropped images, which a crop keeps unless you pressed Keep crop or switched the safeguard off. Without them, the button is dark and the section explains why; re-import the images to align again. The Crop chapter has the details of what a crop keeps.

Why. Alignment is run on the developed, neutral images, never on the camera’s JPEG previews, so that what it measures is what Stack will use. That is also why the button waits for every image to develop.

Show aligned images

A tick box beside the button, with the sub-caption Display registered result in viewer. With it on, the viewer shows each image as it sits after alignment: shifted, scaled and turned onto the reference, so that a detail stays in one place as you step through the images. With it off, the viewer shows the original frames.

Toggle it to see what the alignment did. The reference image looks the same either way, since the others are warped onto it. The program turns the box on whenever you enter Align or Crop, and again when a run lands; its only job is to let you compare.

In the viewer, drag to pan and scroll to zoom. Click an image in the list, or step through them with the same keys as on Images, and watch an edge near a corner of the picture: on a well-aligned stack it stays put.

The last run

The banner under the button reports the run. Before a run it carries the step’s purpose and a dash in every row; after one it reads “{n} images adjusted.” with Alignment completed at the right, and five facts under it.

A run with no fallbacks, a monotone trend and no outliers needs nothing more from you: go to Crop. After an import the program goes there by itself when the run lands; Settings › General › Go to Crop after aligning keeps it on Align instead. Any of the three in the warning colour is worth a look in Diagnostics before you spend time on a stack.

Why the scale trend matters. As focus steps from near to far, magnification changes steadily in one direction. A scale that does not run one way through the stack usually means an image is out of focus order, which the camera can do when it writes a bracket, and which the engines cannot tell from a real change of depth. Diagnostics can sort the stack by the solved scale and align again.

Diagnostics

Before a run the section says “Run Align to see each image’s scale, rotation, shift and residual.” After one it opens by itself, and stays folded afterwards if you fold it. Folded, its caption shows the scale trend’s verdict and a count of the images off it.

The findings come first:

• “Scale changes monotonically through the stack.” or “Scale is not monotone through the stack — an image may be out of focus order.”
• “Off the scale trend: images {numbers}”, if any. An image is off the trend when its scale departs from the straight line through the others by clearly more than the rest scatter, and by at least 0.2 per cent.
• “{n} image(s) fell back to the chained estimate.”, if any.

Reorder by scale

A button in the findings row, available only when the trend is not monotone: “Sort the stack by solved magnification, then align again.” It puts the images in the order their magnification says they were focused, keeps the reference on the same image, and runs the alignment again from that order. Stack results from the old order are dropped. The new order shows on Images as well. This is the fix for a bracket the camera wrote out of sequence; it is not for a stack whose scale wanders because the alignment itself went wrong, which the fallbacks and residuals show.

The table

One row per image, scrolling in both directions when it is wider or taller than the panel.

The how column reads one of:

Rows off the scale trend and rows that fell back are painted in the warning colour, and an image that fell back also carries the ALIGN badge on Images. Step to it in the viewer with Show aligned images on and look at an edge: if it sits where the neighbours’ edges sit, the fallback did no harm. If it does not, the usual causes are a frame that does not belong in the bracket, or a preset that does not match the capture.

Tip. On a rig, the scale column should walk steadily from one end of the stack to the other and rot should stay near zero. A row where both jump is a frame to look at on Images, where the scan on Images › Curate may already have flagged it.

Status messages on this step

The status bar reads “Drag to pan · scroll to zoom · toggle Show aligned to compare”. At its right the latest message shows: “Aligning” while the run is going, “Alignment complete” when it lands, and “Undoing the crop, then aligning” when Align is pressed on a baked crop.

The step card reads “{n} images aligned” once a run is in and no crop has been baked yet. While the run goes it shows the count and the time left, with Pause (“Hold at the next image; Resume carries on.”) and Stop (“Stop at the next image; nothing already finished is lost.”).

A dark step card says “Import at least two images first”. A step card reading “Skipped” means a later step went ahead without an alignment; hover it: “Skipped — the images are stacked unaligned; Align first for a sharper stack”. That is only right for images that were registered before import, which you tell Stack with Images are already aligned.

The Alignment section’s own status line reads, in order of what the program finds: “Warps every image onto the reference so focus is the only thing that changes through the stack.” before a run; “{n} images adjusted.” after one; “Baked — Align puts the full-size images back first, then aligns them.” when a crop is baked and the uncropped images are kept; and “Baked, and the originals were freed — re-import the images to align again.” when they are not.

The factory shortcut Go to Align is Cmd Shift 3; the Reference chapter lists it with the others, and Settings › Shortcuts rebinds it.