PROMPT GUIDE

How to Fix Bad Hands in AI Art: The Complete Guide

Hands are the eternal struggle of AI art, right? Rinka has seen hundreds of pictures with six fingers... But don't worry: with a four-layer approach of negative prompts, composition, poses and fixing afterwards, you can cut broken hands way down. Today I'm putting it all together in one place.

Last updated: 2026-07-10 approx. 3 min read

日本語版

Why is AI so bad at drawing hands?

Hands have lots of joints, and they look very different depending on the angle. Hands also appear in every possible way across the training data, so the AI struggles to consistently reproduce "how many fingers there are and how they bend." That's exactly why there's no single fix: the right answer is to combine negative prompts, composition, pose tags and fixing afterwards.

Start by building a base with your negative prompt

The first step for fixing hands is the negative prompt. Adding tags that describe wrong finger counts or shapes lowers how often hands break. Overdoing it makes the image stiff, so keep the weight at around 1.2 at most.

Go-to negative tags for hands

Tag	Meaning	Note
`bad hands`	Suppresses broken hands in general	The basic hand negative. Start here
`extra digits`	Suppresses too many fingers	The classic fix for six fingers
`fewer digits`	Suppresses too few fingers	More stable when paired with extra digits
`extra fingers`	Suppresses extra fingers	Works much like extra digits. Fine to use both
`missing fingers`	Suppresses missing fingers	For when fingers vanish in gripping poses
`bad anatomy`	Suppresses broken anatomy overall	Insurance that helps all joints, not just hands
`extra arms`	Suppresses extra arms	Good to include when you specify complex poses
`(bad hands:1.2)`	Slightly strengthens the hand fix	Above 1.3, poses tend to get monotonous, so be careful

Negative example: worst quality, low quality, (bad hands:1.2), extra digits, fewer digits, missing fingers, bad anatomy, extra arms, blurry

Flip your thinking: composition tricks that hide hands

Honestly, the surest fix is to not make the AI draw hands at all. Put them in pockets, clasp them behind the back, or give the character something to hold. Tags like these reduce how much hand detail there is, so there's simply less room for things to break. Rinka always uses this trick for an important shot.

Pose tags that naturally hide or tidy up hands

Tag	Meaning	Note
`hands in pockets`	Both hands in pockets	Great with casual outfits. Hands disappear completely
`arms behind back`	Both arms behind the back	A standing-pose classic. Also gives a neat, demure feel Details
`hands behind head`	Both hands behind the head	A cheerful pose. Fingers are mostly hidden by hair
`holding cup`	Holding a cup	With holding tags, fingers follow the object and stay stable
`holding book`	Holding a book	Hand shape settles easily in reading scenes
`hand on own cheek`	Hand on own cheek	Only the fingertips show, so it rarely breaks Details
`own hands together`	Own hands put together	Praying or pleading pose. The hands support each other and stay stable
`crossed arms`	Arms crossed	Hands are hidden by the arms. Also looks dignified Details
`hand in own hair`	Hand touching own hair	Fingers blend into the hair and look natural

masterpiece, best quality, 1girl, solo, long hair, brown eyes, casual, jacket, hands in pockets, cowboy shot, street, evening, smile, looking at viewer

Framing that makes hands small or puts them off-screen

Broken hands stand out more the bigger the hands appear. Put the other way around, if you pull the camera back and make hands small, minor glitches stop being noticeable. Use framing tags to control how hands show up.

wide shot / full body: pulled back, hands get small and glitches don't stand out
upper body + arms behind back: in an upper-body shot, you can move the hands off-screen
portrait / face focus: when the face is the focus, hands don't appear in the first place
from behind: from behind, you only see the backs of the hands
dutch angle: the eye follows the tilted composition, so less attention goes to the hands

Does (perfect hands:1.2) work? Combine it with fixed poses

You've probably seen people put perfect hands or beautiful hands in the positive prompt. To be honest, the effect depends on the model, and you shouldn't expect dramatic improvement. Specifying "what the hands are doing" with concrete pose tags works much better. The vaguer the instruction, the more the AI breaks, so lock the action to one thing, like waving, peace sign or pointing. That said, peace signs and pointing make finger shape matter, so they're harder; if they fail, switching to a hiding pose is the smart call.

masterpiece, best quality, 1girl, solo, ponytail, school uniform, holding book, sitting, library, indoors, soft lighting, upper body, gentle smile, depth of field

Rinka's tips

Rinka's way: the bigger the hands appear, the higher the failure rate, so key images are designed with "holding" poses or "hiding" poses
Piling on negatives makes every pose look the same, so I keep hand-related negatives to 4-5 at most
Generating 10 images and picking the one with good hands (the "numbers game") is a perfectly valid fix too. Fixing the seed to keep the composition is another option

The last line of defense: redraw just the hands with inpainting

If the hands still come out broken, use the inpaint feature in img2img to mask only the hands and redraw them. The tips: make the mask a bit larger than the hand, narrow the prompt down to what the hand is doing (holding cup, etc.), and start with a Denoising strength of around 0.4-0.6. With an auto-fix extension that detects hands, like ADetailer, the hands get redrawn right as the image is generated, which is even easier.

Fixing hands takes both prevention (negative prompt + composition) and treatment (inpainting). You don't need to do it all at once; try starting with four negative tags and one hiding pose.

FAQ

Frequently Asked Questions

How many hand-fix tags should I put in the negative prompt?

I recommend a basic set of four: bad hands, extra digits, fewer digits and bad anatomy. Adding too many has the side effect of making poses monotonous, so keep it to around five.

Is there any point putting perfect hands in the positive prompt?

How well it works varies a lot by model, so don't rely on it too much. Locking in the hand action concretely, like holding cup or hands in pockets, works more reliably.

Can I fix only the hands after generating?

Yes. Use inpainting to mask a generous area around the hands, then regenerate with a prompt narrowed down to the hand action tags. Auto-detection extensions like ADetailer are handy because they fix hands during generation.