Six fingers. A thumb where no thumb should be. Two hands fused into one. Bad hands are the #2 "obviously AI" tell after plastic skin — and in NSFW images, where hands are often front and center, they're even harder to hide. Here's why it happens and what actually fixes it.

Why diffusion models break hands

It's not random. Three things stack against hands:

Hands are tiny in pixel terms. At 1024px, a hand might be 60×60 pixels. The model simply doesn't have enough resolution to "decide" five distinct fingers, so it guesses — and guesses wrong.
Hands are high-variance. Faces are always roughly the same layout (two eyes, nose, mouth). Hands appear in thousands of poses, angles, and degrees of occlusion. The training signal is much noisier.
The model averages. When unsure, diffusion blends plausible options — and the blend of "four fingers" and "five fingers" is six.

So the real cause is resolution and variance, which tells you what does and doesn't fix it.

What does NOT work

Let's kill the myths first:

❌ "perfect hands, detailed hands, five fingers" in the prompt. Counting words don't make the model count. This rarely helps and sometimes makes it worse (it draws attention to hands without giving it more pixels).
❌ Stacking anatomy words. "anatomically correct, perfect anatomy, detailed fingers" is magic-word stacking — see the prompt traps post.
❌ Just rerolling forever. It works eventually, but it's the slowest path.

What actually helps

In rough order of effectiveness:

1. Negative prompt (cheap, real effect)

This is the highest-leverage prompt change. Add the standard hand-failure terms to your negative:

extra fingers, fused fingers, too many fingers, missing fingers,
mutated hands, malformed hands, extra limbs, deformed hands, bad hands

On ximages this set is auto-applied, so you don't need to type it.

2. Compose hands away from the failure zone

You have more control than you think:

Hide or simplify hands in the pose: hands relaxed, partially out of frame, holding something, behind the back, or resting flat. A fist or a hand gripping an object has far fewer ways to go wrong than five splayed fingers.
Prompt the pose explicitly: hands resting on lap, one hand on hip, hand holding a cup. A defined, simple pose beats leaving it to chance.

3. Let an automatic hand detailer re-render them

The actual fix for the root cause (too few pixels) is to re-render the hand region at higher resolution. This is what a "hand detailer" pass does: it detects the hands, crops in, regenerates them with more detail, and blends back. ximages runs this automatically alongside the eye/face detailer, which removes most finger failures before you ever see them.

4. Inpaint as the last resort

When one image is great except for a hand, don't reroll the whole thing — inpaint just the hand. Mask the bad hand, regenerate only that region (optionally with a low denoise so it stays consistent), repeat until it's right. This is the manual version of what the detailer automates.

The realistic workflow

Negative prompt has the hand-failure terms (auto on ximages)
Pose hands simply — gripping, relaxed, or partly out of frame
Auto hand-detailer catches most of what's left
Inpaint the rare survivor instead of rerolling the whole image

Don't forget the same logic applies to feet and teeth

Feet (especially toes) and teeth fail for the exact same reason — small pixel area, high variance. The same playbook works: negative-prompt the failure terms, compose them simply, and detail/inpaint rather than reroll.

Try it on the generator — the automatic hand and face detailing is on by default, so most of this is handled before the image reaches you. For the #1 tell (skin), pair this with the plastic-skin fix.

How to fix bad hands in AI NSFW images · ximages