Star Reduction and Starless Processing: What It Is For

Two images of the same nebula, same integration, same telescope. One looks deep and structured. The other looks like a field of stars with something faint behind it.

The difference is usually not the data. It is that one of them was processed as two images and recombined.

The conflict

Stars and nebulosity want opposite things from a stretch.

Nebulosity is faint and low-contrast, and it needs an aggressive non-linear stretch to become visible at all — lifting the dim end hard while holding the bright end back.

Stars are the brightest things in the frame and already visible. Apply that same aggressive stretch and every star bloats. Cores clip to white, halos spread, and a field that had three hundred distinct points becomes a wash of overlapping blobs. Meanwhile the nebula you were stretching for is now competing with them for attention.

The tighter the star field, the worse this gets. Anything in the plane of the Milky Way — Sagittarius, Cygnus, Carina — is dense enough that the stars can genuinely bury the target.

You cannot resolve this with one stretch, because there is no curve that is aggressive at the faint end and gentle at the bright end in a way that treats a nebula’s dim outskirts differently from a star’s dim outskirts. They are the same brightness. The difference is structural, not tonal.

So: separate them, treat each properly, put them back.

How separation works

A star is small, round, and has a steep brightness profile. Nebulosity is large, irregular, and changes gradually. Those are different spatial scales, and that is the handle.

Two broad approaches:

Scale-based. Decompose the image into spatial frequencies, identify the components that behave like stars, and remove them. This is deterministic and inspectable — you can see exactly what it did and why. It struggles where a star sits on top of bright structure, and it can leave small dark pits where a star was.

Learned. A neural network trained on many star/starless pairs, which infers what belongs behind each star and paints it in. StarNet and Starless AI are the well-known ones. Noticeably cleaner, particularly in crowded fields. The trade is that it is inventing plausible background rather than measuring it — usually harmless, occasionally not, and worth knowing when you present the result as data.

Akastroid’s starless split is the first kind. It suppresses stars using the image’s own coarse scales, which is honest about what it is doing and is not as clean as a trained model in a dense field. That limitation is stated in the README’s list of limits rather than buried, and the step sits behind a single function so it can be replaced when a good offline model is available.

What to do once they are apart

On the starless layer: stretch as hard as the data allows. Push contrast into the faint structure. Denoise properly — with the stars gone, a denoiser stops mistaking small bright points for detail to preserve, and gets noticeably better. Sharpen the nebula without sharpening star edges into rings.

On the star layer: stretch gently. Keep cores unclipped so colour survives — a clipped star is white, and white stars are the giveaway of an over-processed image. Reduce their size if you want to.

Recombine, usually by screen or lighten blending, adjusting how strongly the stars come back.

That last control is the actual point of the exercise. You are choosing how prominent the star field is, rather than accepting whatever a single stretch handed you.

Star reduction without going starless

You do not always need the full split. Reduction alone — shrinking stars in place — handles many images and is less invasive.

Morphological approaches erode the star profile slightly, pulling in the outer pixels while leaving the core. Done gently this is nearly invisible and buys real breathing room.

Done heavily it is obvious and unpleasant: dark rings around stars, or stars that look punched into the image rather than shining through it. Reduction is a step where twice as much is not twice as good.

When not to do this at all

Star reduction has become close to automatic in some processing circles, and it should not be.

Open clusters. The stars are the subject. The Pleiades, the Double Cluster, M35 — reducing the stars removes the target.

Globular clusters. Same, more so.

Wide-field Milky Way. The star field is the structure. A starless Milky Way shot is a picture of dust lanes with the galaxy taken out.

Any image where the star field is doing compositional work. A small nebula against a rich field can be far more striking than the same nebula isolated.

The technique is for when stars are competing with your subject. When they are the subject, or part of it, leave them.

There is also a matter of taste worth naming. Heavily star-reduced narrowband has become a recognisable house style — smooth, saturated nebulosity with a scatter of tiny points — and some people find it looks synthetic. Neither view is wrong. But it is a choice being made, and it is better made deliberately than by following a default.

Where it matters most

Narrowband. Stars come out of a narrowband stack small, magenta and structurally odd, because the filter blocked most of their broadband light — the mechanics of that are in narrowband filters explained. Those stars are rarely worth keeping as captured.

The standard workflow: process the nebula from the narrowband data, take the stars from short broadband exposures shot separately, and combine. Two entirely different acquisitions serving two entirely different parts of one picture, which is only possible because they are separated in the first place.

The short version

• Stars and nebulosity need opposite stretches. Separating them lets each have its own.
• Scale-based separation is inspectable; learned separation is cleaner and invents background.
• The real gain is control over how prominent the star field is.
• Denoising and sharpening both work better with stars removed.
• Do not do it to clusters or wide-field Milky Way. The stars are the subject there.
• Gentle reduction beats aggressive removal almost every time.