Why AI Images Get Hands and Fingers Wrong (and How to Fix It)

AI image models get hands wrong because they were trained to reproduce how things look, not how they work — and a hand is a set of fingers whose correct count and joints only make sense if you know anatomy. The model sees plausible finger-like shapes and stops. The fix is to give it less hand to get wrong, or more references to copy from.

Disclosure: VdoBloom is our platform.

Why it happens

Most images in a model's training data are photographs where a hand occupies a small part of the frame. So the model learns a rough impression of a hand rather than the structure: five fingers in some arrangement. When a hand is large, close to the camera, or makes a specific gesture, there is not enough learned structure to fill in, and you get extra fingers, merged knuckles or a thumb in the wrong place.

There is a second reason. Hands are almost always a small part of a training photo, and they are usually partly hidden, holding something, or moving into a blur. So the model sees far fewer clean, high-detail examples of a hand than it sees of a face. A face is the subject of millions of portraits; a hand is a supporting detail in millions of pictures. The model is simply better practised at the thing it saw more of.

The specific ways hands fail

Knowing the shape of the error helps you spot which fix to reach for, because different failures come from different gaps in the model's understanding.

Wrong finger count. Six fingers, or four, usually on the hand closest to the camera. The model is filling a region it cannot fully resolve.
Merged or webbed fingers. Two fingers that share one outline, common when the hand is small and the fingers are close together.
Thumb in the wrong place. A thumb appearing on the far side, or a second thumb appearing near the little finger.
Joints bending backward. A finger that folds toward the back of the hand, or a knuckle that bends both ways at once.
Stretched anatomy. Fingers longer than the palm, or a hand that is too large for the wrist it attaches to.
Bad hand-to-wrist join. The hand looks correct but floats slightly, disconnected from the arm.

What actually helps

Make the hands smaller. A hand further from the camera, or partly out of frame, has fewer pixels to get wrong. If the pose does not need detailed hands, keep them low, behind something, or in a pocket.
Describe the pose simply and specifically. “Both hands relaxed at her sides” is easier than “hands intricately interlaced”. Specific, simple poses give the model less to invent.
Generate a few and pick. Hands are the most variable part of an output; a bad hand in one generation is often clean in the next for the same prompt. Run the same prompt a couple of times.
Use a reference image. Tools that accept a reference — Image Edit and the reference-capable generators — do better than pure text, because the hand shape can be copied from the input.
Fix the hand, not the image. Take a good result with a bad hand into Image Edit and ask for the hand to be corrected, rather than regenerating the whole picture and losing a good face.

How to describe a hand in a prompt

If the pose genuinely needs visible hands, the wording matters more than usual. A few habits help:

Name the position, not the anatomical detail. “One hand resting flat on the table” is a pose the model can draw. “Fingers slightly curled with the index extended” is a construction it must guess.
Use poses with a reference object. A hand holding a mug, a phone or a rail gives the model a familiar shape to anchor to, and the object hides the parts it would otherwise have to invent.
Avoid asymmetric gestures. “Peace sign with the left hand only” is harder than “both hands in fists”. Symmetry and simple shapes are safer.
Say where the hands are relative to the body. “Hands behind her back”, “hands in her pockets”, “arms crossed” — all of these remove the hand from the hardest part of the frame.

Fixing a hand after the fact

When the rest of the image is exactly what you want and only the hand is wrong, do not throw the whole thing away. Take the finished image into Image Edit and ask for the specific correction rather than a full regeneration:

Describe the result you want, not the problem. Say “both hands relaxed, five fingers on each, resting on the table” rather than “fix the extra finger”.
Keep the rest of the scene out of the instruction. If you mention the face or the background, the model may redraw those too and you lose what you liked.
If the first correction is close, correct it again rather than starting from scratch. Two or three small edits usually land better than one big regeneration.

Checking your result properly

Bad hands survive review because people look at the whole picture at thumbnail size. Before you publish, zoom the image to full size and check the hands the way an editor would:

Count the fingers on the hand nearest the camera. That is the one most likely to be wrong.
Trace the thumb. Is it on the correct side? Is there only one?
Follow each finger to its knuckle. Do the joints bend the way a real hand bends?
Check where the hand meets the wrist and the arm. Does it connect cleanly?

Common mistakes

Adding the word “hands” and expecting a fix. Naming the problem does not teach the model anatomy. Changing the pose or the framing does.
Regenerating everything. A single good generation with one bad hand is usually better salvaged in Image Edit than rerolled from a blank prompt.
Judging at thumbnail size. The hand often looks fine small and falls apart at full size.
Asking for an intricate gesture. The more specific and complicated the hand pose, the more the model has to invent.

Which models do better

Newer models handle hands noticeably better than older ones — it is one of the things each generation improves. On VdoBloom, try a current model like Nano Banana 2.1 or Nano Banana Pro for images where hands are prominent, and compare a couple of models on the same prompt.

Frequently Asked Questions

Why do extra fingers appear?

The model reproduces how hands look, not how anatomy works, so when it has too little structure to work from it improvises — often adding a finger.

Will a better prompt fix hands?

Sometimes. Reducing how prominent the hands are, and simplifying the pose, helps more than adding the word “hands” to the prompt.

Can I fix only the hand?

Yes — use Image Edit to correct the hand on a result you otherwise like.

Are hands always going to be a problem?

Less so each generation, and easily avoided by posing or framing them out. Treat visibly detailed hands as the hardest part of any AI image and plan around them.

Does a reference image really help?

Usually yes, because the model can copy a real hand shape from the reference instead of inventing one. It is the most reliable fix when the hands must be large and detailed.

Should I hide the hands to be safe?

If the pose allows it, yes. Pockets, behind the back, holding an object, or cropping below the wrist remove the problem entirely, and often look more natural than a hand the model had to guess.

Questions people ask about this

Why does AI make hands look weird in the first place?

Because the model was trained to reproduce how things look, not how they work. A hand only looks right if the finger count and joints make sense, and the model does not know anatomy — it just sees plausible finger-like shapes and stops. When a hand is large or makes a specific gesture, there is not enough learned structure to fill in, so it improvises.

Is it because there isn't enough training data?

Partly, but not in the way people usually mean. There is plenty of data — the problem is that hands are almost always a small, partly hidden or blurred part of a training photo, so the model sees far fewer clean, high-detail examples of a hand than of a face. It is simply better practised at the thing it saw more of.

Why do AI images get the number of fingers wrong?

The model is filling a region it cannot fully resolve, so it adds or drops a finger rather than counting. This shows up most on the hand closest to the camera, where there are more pixels to get wrong. Zooming to full size and counting the fingers on that hand is the quickest way to catch it.

Why do artists and AI both struggle with hands?

Hands are a set of fingers whose correct count and joints only make sense if you know anatomy, so they are the hardest part of any figure to get right. This guide treats visibly detailed hands as the hardest part of any AI image for the same reason. Planning around them — posing or framing them out — is easier than fighting them.

Why do AI images of people have obvious flaws like wrong fingers?

Because people review the whole picture at thumbnail size, where a bad hand often looks fine. The flaws survive review rather than being created by it. Checking the hands the way an editor would — counting fingers, tracing the thumb, following each finger to its knuckle — catches most of them before you publish.

Can I just ask the AI to fix the hands?

Naming the problem does not teach the model anatomy, so adding the word “hands” to a prompt rarely fixes anything on its own. What helps is changing the pose or the framing, or correcting the hand on a result you otherwise like. Describe the result you want rather than the problem you see.

→ Generate an image · Create a free account