Planetary stacking for Mac, Windows and Linux

No neural network, and a sharpener that can say no

Akastroid runs no neural network and ships no model file. Instead the sharpening stage tries a ladder of strengths, measures what each one did, and takes the first that passes every guard. When none does, it writes which guard stopped it to the processing log. These are the exact messages:

• "No restoration: a 5.1 px blur was measured, but correcting it recovered no detail this stack did not already have."
• "No restoration: the limb rises over 6.2 px on a body 240 px across, which is the planet's own limb darkening rather than a blur that can be undone."
• "No restoration: 240 frames is too thin a stack to sharpen a 4.8 px blur without amplifying grain faster than detail."
• "No restoration: a 3.9 px blur was measured, but every correction strong enough to matter also rang at the limb or amplified the sky."

The first of those is from a real Moon capture in our regression set, recorded against every release. An AI sharpener given the same stack can invent a plausible festoon, and you would have no way to tell. That is why this is a guard and not a slider.

A planet that turns

Jupiter rotates fast enough to smear a long capture. Usually that is a problem, and an animation of it has meant WinJUPOS. Akastroid stacks the capture in time slices instead, each slice sharing one reference frame and one set of sharpening strengths, and writes the result as a GIF or an AVI. The planet turns rather than flickers, and it comes from the same video you already shot.

Measured against AutoStakkert

AutoStakkert is the free standard and most planetary images you have admired went through it. Stacking the same eight captures in both, unsharpened, compared at matched physical frequencies with the noise floor measured in the same tile of both images: the two carry the same detail, and ours is quieter on every capture. It was behind on red and blue for one-shot-color captures (a demosaic problem, not a stacking one) until each colour began to be integrated from its own sensor sites in September 2026; on the Mars capture every channel now reads 1.03x AutoStakkert's on the unsharpened stack.

We used to claim sharper on every capture. That was withdrawn in September 2026: the rulers behind it carried a contrast confound and a noise floor that subtracted more from whichever image resolved more, which penalised the sharper image. The method, and both retractions, are on the AutoStakkert comparison.

So the reason to use Akastroid is not the stacker. AutoStakkert stops at the stack; grading, stacking, dispersion correction and wavelet sharpening happen here in one pass.

The Mac situation, plainly

There is no native macOS build of AutoStakkert, RegiStax, PIPP or WinJUPOS, and there never has been. They are Windows programs written by people solving their own problem on Windows, several of them well over a decade ago, and they are free. None of that is a complaint, but it means the planetary advice you will be given, on every forum and in every tutorial, assumes a machine you may not own.

The usual answers are Wine, CrossOver, Parallels or a spare Windows laptop. They work, more or less, and they are a second computer to maintain in order to sharpen a video of Saturn.

One native exception is worth knowing about: Lynkeos is free, open source and has aligned and stacked planetary video on the Mac for close to twenty years. If you want a free stacker and are willing to drive it yourself, it is a good one.

It does not do the whole job. There is no atmospheric dispersion correction, the sharpening is yours to dial, and it does not do the Moon or deep sky, so you need a second program for the rest and a third for the other half of the hobby. The guide to what runs on macOS, and what does not has the full list, including the free tools. Siril and PixInsight are both native on Apple silicon, but neither does planetary work, and both take a lot of study before a first image.

Why planetary work is a different problem

Deep sky imaging fights photon noise: the signal is faint, so you integrate for hours to accumulate it. Planetary imaging has the opposite problem. Jupiter is bright enough to expose in milliseconds, and what ruins the picture is the atmosphere: several kilometers of moving air between the planet and the sensor, smearing the detail differently in every frame.

The answer is lucky imaging. Record thousands of very short frames, and the turbulence will have been briefly kind in a few hundred of them. Find those, align them, average them and sharpen the result. Nothing from a deep sky workflow transfers: there are no stars to register on, quality means sharpness rather than star shape, and the final image lives or dies on the sharpening rather than on the integration.

Grading: finding the frames the seeing left alone

Every frame in the capture is scored, and the measurement is more careful than it looks. The obvious approach, Laplacian energy (the standard focus measure), gets planetary frames backwards, because a Laplacian responds most strongly to the highest spatial frequency present, and in a short planetary exposure that is read noise. Score frames that way and the noisiest frame wins.

So the measurement is the structure at the scales a planet's detail occupies, divided by the noise in that same frame, and it is restricted to the region around the disc so that empty sky cannot dilute it. What comes out is a ranking that matches what your eye would pick if you had the patience to flick through seventeen thousand frames.

You choose what percentage to keep. A graph of the whole capture shows how the seeing changed while you were recording, which is worth reading on its own: it tells you whether to record for longer next time, or to wait an hour for the planet to climb.

Multi-point alignment

Aligning frames on the planet as one rigid object is not enough. The atmosphere does not only shift the disc, it warps it: the equatorial belt moves one way while the polar region moves another, within the same frame. Stack that as a rigid object and you get a sharp limb and a mushy middle, and no amount of sharpening recovers what the misalignment averaged away.

Akastroid tracks the disc in pieces on a grid chosen from the disc size, so the local warping comes out before anything is averaged. More than any other setting, this decides whether a planetary stack is soft or sharp, and it is the reason a capture can look disappointing in one program and detailed in another with identical source data.

The surviving frames are then averaged, not sigma clipped. Once the bad frames have been culled the survivors differ only by noise, and a plain mean extracts the most signal from them. Clipping here would throw away good data to solve a problem that grading has already solved.

If your planet is undersampled, as most are at a few hundred pixels across, the integration is done onto a finer grid, recovering detail below the original pixel scale from the sub-pixel shifts between frames.

Dispersion correction

The atmosphere refracts blue light more than red, so a planet low in the sky arrives on the sensor as three slightly separated images. It shows as a blue fringe on one edge and a red fringe on the other, and it is the most common reason an otherwise good capture looks wrong.

Red and blue are realigned to green before sharpening. Done afterwards, sharpening has already amplified the fringes into something no alignment will remove cleanly.

Sharpening by scale

The stack is sharpened per scale: six sliders from finest to coarsest, each with its own noise suppression, so you can lift the fine detail in the cloud belts without amplifying the grain that lives at the same frequency. Moving a slider re-applies the sharpening without re-stacking.

There is also deconvolution, which measures the blur from the planet's own limb and partly undoes it. It is applied before sharpening and it is deliberately conservative: deconvolution invents plausible detail when it is pushed, and a planet has no star in frame to check the answer against, which is why the control starts low and the guidance on it says to raise it slowly and compare.

The Moon

Lunar captures take the same path. The Moon's advantage is that it is bright and full of contrast, so grading has plenty to work with. The difficulty is that it fills the frame, and seeing varies measurably across a field that wide, so multi-point alignment does more here than anywhere else.

The Moon has its own page: lunar imaging, where multi-point alignment across a surface rather than a disc does the work. As photography rather than as software it is covered in the guide on photographing the Moon.

What this replaces, and what it does not

Akastroid covers the common path (grade, select, align, stack, sharpen) in a single window, with the capture, the grading graph and the sharpening sliders all on screen at once, so you can go back and forth between them without exporting anything.

AutoStakkert and RegiStax are free, mature and very good at what they do, and there are corners (unusual derotation workflows, unusual capture formats, specific manual controls) where they go further. Somebody already fluent in that chain has no particular reason to change.