AI Artefacts: Hands, Teeth, Jewellery and Other Weak Spots

The errors of generative models are not evenly distributed. There are places where they appear almost always, and knowing the reason lets you find them in seconds.

Hands

The famous case. The cause is not the complexity of the shape but its variability: a hand takes thousands of positions, fingers overlap one another, and in training data hands are usually small and indistinct. The model learned that «several elongated objects go here» but not that there are exactly five.

Teeth

In photographs teeth occupy few pixels and are often soft. The model reproduces «a light area in the mouth» without understanding the boundaries between individual teeth. In a close smile this shows immediately.

Jewellery and repeating elements

A chain, a bracelet, lace, lacing — anything made of identical links. The model grasps the texture but not the count: links merge or double, and a chain can disappear under skin and never re-emerge.

Text

Lettering on clothing and objects turns into pseudo-letters. Same reason — the model reproduces the appearance of text rather than its meaning.

How to search quickly

• Hands and fingers first
• Then boundaries: where an object crosses the body
• Then symmetrical pairs: earrings, shoes, sleeves — they often differ
• Then small repeats: teeth, links, eyelashes
• The background last: errors there are least visible

What can be repaired

Regenerating a selected region locally solves most problems with hands and jewellery. What cannot be fixed is a compositional error — when the placement of the figure itself is wrong: regenerating is easier than repairing.

This material is intended for adults over 18 and is informational in nature.