Skip to content

Astronomy and space

Space is the subject AI image models render most beautifully and least accurately, producing images that look like astrophotography while showing planets, scales and colours that do not exist.

There is more space imagery in the training data than almost any other scientific subject, so the output is spectacular. It is also confidently wrong about everything measurable: ring systems on the wrong planets, moons at impossible distances, nebulae in colours nothing emits. As decoration it is unmatched.

What this cannot do

Never present a generated space image as a real object. Saturn will have the wrong number of rings, Mars the wrong colour, and any "photograph" of a named galaxy is a fabrication. NASA and ESA both release genuine imagery into the public domain — for anything real, use theirs.

4 prompts

Deep space nebula

Headers and backgrounds for astronomy content.

Vast nebula of dust and gas in deep space, layered clouds lit from within, scattered foreground stars, rich blues and violets with a warm core, extreme depth, astrophotography aesthetic, high dynamic range

Things to change

  • — "monochrome" for something more restrained
  • — "warm amber and rust" for a different palette entirely

Planet from orbit

Space and planetary science sections. "Unnamed" is deliberate — it stops the model claiming a real planet.

View of an unnamed rocky planet from low orbit, curved horizon catching the light of a distant star, thin atmospheric glow along the limb, deep black space above, subtle surface detail, cinematic composition

Things to change

  • — "ice planet" or "volcanic planet" for variety
  • — "with a small moon in the distance" for scale

Night sky, long exposure

Anything atmospheric. Closest to a real photograph and therefore hardest to get wrong.

Long exposure photograph of the night sky over a dark landscape, dense star field, faint band of the Milky Way, silhouetted horizon along the bottom, no light pollution, natural colour, minimal composition

Things to change

  • — "star trails, circular" for a several-hour exposure look
  • — "with a lone figure for scale"

Orbital diagram, stylised

Section dividers and card images. "Not to scale" in the prompt sets the right expectation in your own head too.

Stylised diagram of a small planetary system, concentric orbital rings around a central star, flat minimal vector style, thin lines on deep space blue, geometric and clean, not to scale, no text or labels

Things to change

  • — "retro 1970s space poster" for a strong period look
  • — "blueprint style"

Common questions

Are AI space images accurate?
No. They are convincing and wrong — incorrect ring systems, impossible moon distances, nebula colours nothing emits. They look like astrophotography, which makes the inaccuracy harder to spot, not easier.
Where do I get real space images?
NASA and ESA both publish extensively, and most NASA imagery is public domain and free for any use including commercial. For a real object, that is strictly better than generating one.
Why is space the model's strongest subject visually?
Volume of training data and forgiving subject matter. There is an enormous amount of space imagery online, and diffuse glowing clouds have no rigid structure to get wrong — unlike a hand, a face or a benzene ring.

Get the numbers right too