Why Is AI Bad at Hands? (And How to Fix AI Hands)

Six fingers, fused knuckles and a hand melting into a coffee cup all trace back to the same few causes, and they explain why AI is bad at hands: a hand can take almost any shape, it's often small or half-hidden in the photos a model learns from, and the model has no skeleton to check its work against. Below is what the research found, whether AI can draw hands now, and how to prompt for a good hand and fix a bad one.

Why is AI bad at hands?

AI is bad at hands because image models learn from huge sets of captioned pictures, and hands are among the hardest things in those pictures to learn. When Stability AI published its SDXL paper in July 2023, the first limitation it listed was human hands. The authors' likely explanation: hands look very different from one photo to the next, so the model struggles to work out their real 3D shape and how far they can bend.

A hand has too many possible shapes

A face nearly always follows one layout, with two eyes above a nose above a mouth, and it usually sits upright. A hand has 19 bones, plus eight more in the wrist, according to the Cleveland Clinic. A 2024 Georgia Tech paper on AI hands puts the problem in numbers: each of a hand's 15 joints can sit at many angles, and hands turn up in almost any orientation, while faces, cars and houses usually appear the right way up. A fist, a wave and a pinch share almost no outline, and the model has to learn each one from flat images.

Four fingers that look alike

Apart from the thumb, the fingers on a hand look very similar. The same Georgia Tech paper suggests this makes it harder for a model to keep count, and to tell a left hand from a right one. Both show up in the paper's own list of common errors: too many fingers or too few, and left hands where a right hand belongs.

Hands are often hidden in the training photos

In everyday photos, hands are tucked into pockets, cut off by the edge of the frame, or wrapped around something, and they often fill only a small corner of the picture. The HandRefiner paper from 2023 says the structure and pose of hands are hard to learn from training images, because hands bend into so many shapes and are so often partly covered. A model that has mostly seen half a hand behind a mug has little to go on when you ask for five fingers in full view.

The model sees pixels, never bones

An image model has no skeleton to check its work against. Diffusion models such as Stable Diffusion start from random noise and clean it away in small steps, guided by patterns from their training images (the guide to how AI image generators work walks through the steps). Nothing in that process counts bones, so a hand that looks right at a glance can still have an extra knuckle. That is the core of why AI is bad at drawing hands in any style, from photos to cartoons.

What AI hand mistakes look like

AI hand errors fall into a few repeat patterns. Knowing them makes a bad hand quicker to spot and easier to describe when you fix it.

• Wrong finger count. Six fingers, or four. The HandRefiner authors name incorrect finger counts as a typical failure.
• Irregular shapes. Fingers that are too long, bend the wrong way, or have odd skin texture.
• Fused fingers. Two or three fingers merged into a mitten shape, often where the fingers touch.
• Extra or missing hands and arms. The Boston University and Adobe study described below labels one figure with three arms and three hands.
• Hands merging into objects. A fork that turns into a finger, or fingers that pass through a phone.
• Mirror mistakes. A thumb on the pinky side, or a left hand attached to a right arm.

Feet go wrong too, for similar reasons. Researchers at Boston University and Adobe Research labeled more than 37,000 images of people from SDXL, DALL-E 2, DALL-E 3 and Midjourney and found that errors cluster at the extremities of the body. Local errors, where one body part looks wrong, showed up most in the hands for all four generators, and hands and feet both came up more often than the torso or legs.

Can AI draw hands now?

Yes, far more often than in 2023, though not every time. Midjourney's April 2025 alpha announcement for V7 said bodies, hands and objects now hold together much better in the fine details. As of September 2026, Midjourney's default is V8.2, which took over on July 24, 2026.

Research points the same way. In the Boston University and Adobe study, DALL-E 3 and Midjourney showed stronger body structure, with fewer errors overall, than SDXL and DALL-E 2. Researchers are also going after hands directly:

• HandRefiner repairs a finished image by repainting only the hand. It rebuilds the hand as a mesh model that always has the right number of fingers, then uses that shape as a guide while the rest of the image stays as it was.
• Annotated Hands, the Georgia Tech paper, adds three extra channels of hand annotations to each training image. Models trained this way produce hands whose finger joints a standard hand-tracking tool recognizes with more confidence.

None of these sources claims hands are solved. Midjourney describes V7's hands as more coherent, and the papers exist because the problem is still open. So the honest answer to why AI can't draw hands properly in 2026 is that newer models get them right much more often, but no maker promises five fingers every time.

How to prompt for better AI hands

Write the hand into your prompt so the model doesn't have to invent the pose. Each habit below leaves something you can check in the result.

• Describe the pose in plain words. Say it the way you'd direct someone posing for a photo: "right hand raised, palm toward the camera, fingers slightly apart".
• Say which hand, and what it holds. "Her left hand holds the leash" is clearer than "holding a leash". Pick objects with a simple shape, like a mug or a book.
• Show fewer hands. A photo of six people has twelve hands to get right. If the hands don't matter, choose a smaller group.
• Skip interlocked fingers. Handshakes, clasped hands and crossed fingers make fingers overlap and hide each other, which is the occlusion problem again. Show the moment just before a handshake instead.
• Give the hand enough of the frame. In a wide shot, a hand may cover a tiny patch of the image. A medium shot or close-up leaves room for knuckles and nails.
• Hide hands on purpose when they don't matter. Pockets, a counter in front, or a large tray held with both hands all keep the fingers out of view.

In Stable Diffusion tools you can also type "extra fingers" into a negative prompt. That steers away from bad hands but doesn't describe a good one, so the pose description matters more. These prompts are written for Toybox AI's Image Generator, and they work as a first draft in any text-to-image tool:

Close-up photo of a woman's right hand holding a white ceramic mug, her index and middle fingers through the handle, thumb resting on top, the other two fingers curled underneath. Relaxed grip, natural knuckles, short unpolished nails. Morning window light from the left, wooden table, shallow depth of field.

Portrait photo of a smiling man in his 40s waving hello, his left hand raised beside his face with the palm toward the camera and all five fingers spread apart. Plain light gray backdrop, soft studio light, eye-level angle, sharp focus on the hand and face.

Medium shot of a baker standing behind a bakery counter with both hands tucked into the front pocket of her flour-dusted apron, smiling at the camera. Shelves of bread loaves behind her, warm light, 50mm lens.

How to fix AI hands after you generate

When everything but a hand looks right, you have five options. Which one fits depends on how much of the image you need to keep.

Selecting an area and regenerating only that part is called inpainting, and the guide to what inpainting is explains how masks work. Midjourney's Vary Region, for example, works on images from version 5 and later: you select the hand with a freehand or rectangle tool, and a smaller selection gives a subtler change. A describe-to-edit tool works without a selection, so name the hand precisely ("her left hand, on the mug") and say what should stay the same.

Troubleshooting AI hands

How to check an AI image for bad hands

Check hands at full size before an image goes anywhere public, because a phone thumbnail hides most errors.

1. Zoom in and count fingers on each hand you can see, background hands included.
2. Match each hand to an arm, and count hands against people.
3. Check the thumb: it should sit on the side nearest the index finger, and the palm should face the way the pose implies.
4. Look where skin meets an object, such as a cup handle, a phone edge or a strap.
5. Check the feet and any lettering too, since text fails for related reasons (see why AI messes up text in images).

Hands are one clue among several when you're judging whether a picture was made with AI, and the guide on how to tell if an image is AI covers the rest. For skin, light and lens tips that make the whole picture more convincing, read how to make AI images look real.

Fix one hand in Smart Editor

When one hand is the only problem, fix it with a written instruction instead of a brush. Smart Editor runs on OpenAI's GPT Image 2.5 and needs a Pro plan.

1. Make the image in Image Generator with the pose written out, then zoom in and find the wrong hand.
2. On a Pro plan, type the fix into the "Want to change something?" box once the result appears, or choose "Open in Smart Editor". You can also upload a saved image to Smart Editor (JPG or PNG, up to 10 MB).
3. Name the hand and the fix. For example: "Make her left hand show five relaxed fingers wrapped around the mug handle, thumb on top. Keep her face, hair, clothes and the background the same."
4. In Smart Editor, tap Apply edit and wait about 30 seconds for the new version.
5. Compare the face, the other hand and any text with the previous version. Each edit redraws the whole image, so small details can shift.
6. To fix something else, keep "Editing" on "Latest result" and type the next change, or switch to "Original photo" to try again from the first version.

An HD edit in Smart Editor and a fix in the revise box each currently cost 85 credits, and plans currently start at $9.99 a month for 1,000 credits (see current prices). A wrong hand can take more than one try, so describe the grip and where each finger goes, not just "fix the hand".