The short answer
AI draws six fingers not because it glitches. It does exactly what it was trained to do; the task with hands is just objectively hard.
The model does not know a human has five fingers. It has no concept of «fingers», «hand» or «skeleton» at all. It has seen millions of flat pictures and remembered how pixels are usually arranged next to each other. That is all. There is no understanding of anatomy inside.
As long as the pixels form a familiar pattern, everything is fine. A hand is exactly the case where the pattern keeps breaking.
Why hands in particular
There are several reasons, and they add up.
Fingers look alike. The model sees a group of elongated pink shapes and adds one more «just in case». It has no way of knowing there are already enough: it does not count, it guesses from the overall look.
Fingers overlap each other. In real photos some fingers are hidden behind others, bent, tucked into a fist. When reconstructing what is hidden, the model may draw an extra one to «fill in» what seems to be missing.
A hand takes thousands of shapes. Compare it with a face: it is almost always frontal and symmetrical, with eyes and nose in their places. The model learned that pattern easily. But a hand can be clenched, open, in profile, holding an object, making a gesture. The variety is such that no clear rule can be derived from it.
The model assembles the picture in parts and does not keep count. It forms the image in fragments and does not hold the global picture in mind. It can add a finger on one side of the hand, forgetting how many there already are on the other.
If you want to understand the mechanics of generation that all this follows from, there is a separate article: how AI image generation works.
The same cause: text, teeth, crowds
Bad hands are the best-known symptom, but not the only one. All these bugs share one root: the model reproduces a pattern without understanding the content.
Text in the image. To the model, letters are ornament. It draws something that looks like letters but does not understand that they form words. The result is pseudo-text: the letter shapes seem right, but you cannot read it. Cyrillic is worse than Latin script because there was several times less Russian text in training.
Teeth. The same case as fingers: many similar identical elements in a row. The model easily adds an extra tooth or makes a smile unnaturally wide.
Faces in the background. A face in focus comes out well, but for people in a crowd behind, it often «melts»: the model pays less attention to them, and the pattern breaks.
Symmetrical objects. Glasses, wheels, building windows. Anything where exact repetition matters can throw the model off.
Understanding this helps predict where the problems will be even before generation: anywhere there are many identical small details in a row.
How to fix it: three levels
Level 1. The right model
The simplest and often the most effective. New, more detailed models handle hands noticeably better than old ones: they were trained on more high-quality images and pay closer attention to fine details.
A practical trick: do not fight one model if it stubbornly ruins hands. Run the same prompt through several models and pick the best result. The same scene comes out differently on different architectures, and often the neighbouring model gets the hands right on the first try.
Level 2. A smart prompt and composition
If the hands are not the main thing in the frame, take them out of focus. Ways to do it:
- hide the hands: in pockets, behind the back, holding an object that partly covers them,
- move them to the background, out of focus,
- choose a shot where the hands are simply not visible (a chest-up portrait with the hands out of frame).
If the model supports a negative prompt, put what should not be there into it: distorted hands, extra fingers, deformed hands. It is not a guarantee, but it improves the odds.
And do not describe the hands in too much detail. «Graceful fingers covered in rings, clutching a wine glass» is asking for a reject: the harder the task for the hand, the higher the chance the model fails it.
Level 3. Inpainting (regenerating an area)
The most reliable way when the picture as a whole worked but a hand let you down.
Inpainting is regenerating part of the image. You select the problem area (the hand) and ask the model to redraw only that, leaving everything else as it is. The model concentrates on a small area, and the chance of getting a normal hand rises sharply.
This is much more effective than generating the whole picture again in the hope of getting lucky this time. Starting over, you risk ruining what is already good, while inpainting touches only the problem.
What not to do
Do not hit the same prompt dozens of times without changes. If the model stubbornly breaks hands in this scene, ten more attempts will not change things. Change the model or the composition, or move on to inpainting.
Do not rely on the negative prompt alone. It helps but does not solve it. «Bad hands» in the negative will not make the model understand anatomy; it only shifts the probabilities slightly.
Do not expect a new model to solve everything. There is progress, but as long as models are statistical rather than understanding, hands will sometimes break. It is easier to build this into your process than to get upset every time.
In short
Bad hands are not a malfunction but a natural consequence of how generation works: the model repeats patterns without understanding anatomy. The same cause produces garbled text, extra teeth and melting faces in a crowd.
There are three levels of cure: choose a newer, stronger model; hide the hands from focus or put the artefacts in a negative prompt; and if the picture as a whole worked, regenerate the problem hand with inpainting. The problem will not go away completely, but it can be made unnoticeable.
