Planetary Sharpening: The Wavelet Sliders, Dispersion, and Why 4× Looks Worse

A planetary stack straight out of the integrator is soft. That is expected: it is the average of a few thousand frames that each wobbled a little differently, and averaging is a blur. The image everyone recognises - belts, festoons, the Cassini Division as a black line - is made in the sharpening stage, and the sharpening stage is six sliders that most people set by copying someone else’s screenshot.

This post is about what those sliders do, what the two related controls (RGB alignment and restoration) do, and about one specific mistake that is easy to make and hard to see: pushing the fine layers to their limit and getting an image that is noisier but not sharper.

What a wavelet layer is

Take the stack. Blur it slightly and subtract the blur from the original; what is left is the finest detail - single-pixel structure. Blur the blur a little more, subtract again; that is the next scale up. Do it six times and you have separated the image into six layers by size: the finest holds pixel-scale texture, the coarsest holds the disc’s overall shading, and the belts and the Cassini Division live somewhere in the middle.

A wavelet slider multiplies one layer before the image is put back together. Set the third layer to 2× and every feature at roughly that size is twice as contrasty. Set the finest layer to 4× and every single-pixel feature is four times as contrasty - and on a planetary stack, most single-pixel features are noise.

That is the entire tool. RegiStax made it famous; Akastroid’s Sharpening panel is the same six layers, Finest to Coarsest, each with a noise-suppression control that runs before the layer is amplified.

Where the starting values come from

They are not a preset. The tuner measures how much fine-scale structure the stack carries relative to its noise, and how deep the stack is, and sets the layers from that.

The depth matters more than it looks. The finest layer of a 60-frame stack is mostly read noise, and amplifying it produces the crunchy, boiling look that says “oversharpened” from across the room. The finest layer of a 2,000-frame stack is different: averaging has beaten the noise down by a factor of forty-five, and at that depth the finest scale holds real detail - festoons, the fine ring gaps, small ovals. The tuner lifts it in proportion, from a floor of “leave it alone” on a thin stack to a modest gain on a deep one, and eases the noise suppression back at the same time so it is not removing what it just amplified.

The values it lands on are a starting point. They are chosen to be safe, which is not the same as best for every capture, and the sliders are there for the difference.

The 4× mistake

Here is a real case. A Saturn capture, 17,934 frames, the best 12% stacked - a good stack from rough air. The tuner set the three fine layers to roughly 1.4, 2.3 and 2.6. The operator, chasing more detail, pushed all three to 4.0, the top of the range, and turned the finest layer’s noise suppression up to compensate.

The result looked softer. Not obviously noisier - the suppression was hiding that - but the Cassini edge was less clean and the belt boundaries less defined than at the starting values, and the globe had a faint grain to it.

What happened: the fine layers on that stack contained detail and noise, in a ratio the tuner had measured. Multiplying both by 4 kept the ratio. Then the noise suppression, turned up to hide the amplified noise, removed detail at the same scale along with it - a threshold cannot tell them apart. The net was less detail and more grain than the starting values, with the extra work invisible because the suppression smoothed the evidence.

The rule of thumb: on a deep stack, the fine layers can go a little past the tuner’s values. If you find yourself raising the noise suppression to make a slider setting tolerable, the slider is too high.

Atmospheric dispersion, and RGB align

A planet low in the sky is a prism’s worth of atmosphere away. Red and blue light refract by different amounts, so the red image of the disc lands slightly above the blue one - a red fringe on one limb, blue on the other. At 30° altitude it is a pixel or two; at 15° it is several, and no amount of sharpening fixes a disc whose colours are in different places.

RGB align measures the offset between the red and blue channels and the green one - on the disc’s structure, after standardising the channels so brightness differences do not fool the matcher - and shifts them into registration. On the Saturn capture above the red channel was 0.63 px off and the blue 0.87 px, in opposite directions: a 1.5-px total split that after correction measured under 0.25 px. That is on by default; you would only turn it off for a capture taken through an atmospheric dispersion corrector, where the offset is already zero.

One thing it cannot do is fix a fringe that is not dispersion. A colour camera with a Bayer filter has its own small channel offsets, and a refractor has chromatic aberration; both shift with focus, not with altitude. RGB align corrects the sum of all of them as one shift, which is right for dispersion and approximately right for the rest.

Restoration

Off by default, and worth understanding why before turning it on.

Restoration is deconvolution: measure how blurred the planet’s limb is, and computationally undo that blur before sharpening. On a deep-sky image the same technique has a built-in check - the stars are point sources, so if the deconvolved stars look wrong, the deconvolution is wrong. A planet has no point source in the frame. The algorithm will produce plausible-looking detail whether or not that detail was in the data, and there is nothing to compare it against.

So it is a slider that starts at zero. Raise it a little, compare with the stack, and stop when you are no longer sure the new detail is real. That point comes sooner than people expect.

Practical order

1. Crop to the planet, so the sharpening is judged on the disc and not on the sky.
2. Leave the wavelets at the tuned values and look. That is the honest stack.
3. Nudge the second and third layers up if the belts are soft. Go back down if the limb starts to ring.
4. Touch the finest layer last, and only on a deep stack.
5. Check the limbs for colour fringe. If RGB align was on and a fringe remains, it is not dispersion and the sliders will not remove it.
6. Try Restoration at a low value if the stack is deep and the seeing was good. Turn it off if you cannot tell what it changed.

The short version

• Six layers by scale; a slider multiplies one layer. Belts live in the middle layers, noise in the finest.
• The tuned values are measured from the stack’s depth and detail, not copied from a preset.
• Pushing the fine layers to 4× and hiding it with noise suppression loses detail. If you need more suppression, lower the slider.
• RGB align corrects dispersion to a fraction of a pixel; leave it on.
• Restoration invents as readily as it recovers. Start at zero.