EfficientNet-B0 fine-tuned on Galaxy Zoo 2

See morphology
in a single glance.

A neural network trained on 25,000 labeled galaxies sorts each image into one of five morphological types. Every card below is classified live by the model.

Elliptical

Smooth, ellipsoidal swarms of older stars with little gas, no spiral structure.

Spiral

Disk galaxies with curving spiral arms tracing star-forming regions.

Barred Spiral

Spirals with a luminous central bar of stars feeding the inner regions.

Edge-on Disk

Disk galaxies seen edge-on, often revealing a thin dust lane.

Merger

Two or more galaxies caught mid-collision, tidally disturbed.

Live gallery

Each card calls the model on page load. Ground-truth labels are shown so you can see when it gets it right — and when it doesn't.

Under the hood

Performance on the held-out test set, the model that produced it, and what it gets wrong.

Held-out test accuracy

78.9%

1,917 of 2,430 test images classified correctly

Test set is held out from training — these images were never seen during fine-tuning.

Accuracy by class

  • Edge-on Disk93.4%467/500
  • Merger85.6%368/430
  • Elliptical82.6%413/500
  • Barred Spiral69.4%347/500
  • Spiral64.4%322/500

Architecture

EfficientNet-B0

4.0M params

Training images

19,445

Galaxy Zoo 2

Epochs

8

AdamW + cosine LR

Classes

5

morphological types

Honest note

The model's real weakness is telling spirals apart from barred spirals — those two classes share most of their morphology and the bar can be subtle even to human classifiers. Edge-on disks are easy because the silhouette is unmistakable.

Try your own

Drop in any galaxy image. The model returns probabilities across all five classes.

Classify your own galaxy

Drop an image here or click to upload. JPEG, PNG, or WebP up to 8 MB.

Works best with the galaxy centered and roughly filling the frame.

Results will appear here.