Keeping two touching hands apart in AI images

Two hands that touch often go wrong in AI images because, where they overlap, the model can no longer tell where one hand ends and the other begins, so fingers merge or the count jumps. A handshake, a high five or two people holding something together is therefore harder than a single hand on its own. Below is why that happens, and how to make a contact pose believable anyway.

Why do two touching hands merge?

Because the contact hides some fingers behind others and the model loses the depth order. A single hand is already hard: a human hand has around 27 degrees of freedom (directions of movement across the finger, thumb and wrist joints), so there are many possible poses that all have to look right. Put two hands against each other and occlusion is added on top: where the hands overlap, the model has to guess which finger sits in front of the other.

That is a separate and harder problem than one hand. In 2020 researchers built a dedicated dataset, InterHand2.6M, of 2.6 million frames precisely because earlier work focused mostly on a single isolated hand (Moon et al., ECCV 2020). A image model also treats "hands" as one concept, and hands often occupy only a few pixels, so the fine detail between two hands is undersampled. You see the result as fused fingers, a wrong count or an impossible grip.

Describe the contact, not just the pose

State exactly which parts touch and which stay visible, instead of only "two people shaking hands". The more concrete you make the contact point, the less the model has to guess.

• Name the count and the side: "two right hands in a firm handshake".
• Describe the grip: "thumbs crossed on top, fingers wrapping around the back of the other hand".
• Say what is visible: "both wrists in frame, fingertips not hidden".
• Keep other hands out of it: mention that the free hands rest at the sides, so a stray third hand does not appear.

This follows the same logic as binding attributes to one subject: the clearer you say what belongs to whom, the less it leaks or merges.

Pick a contact pose with as little overlap as possible

When in doubt, choose the pose where the fewest fingers disappear behind each other, because less occlusion means a more stable result. Not every touch is equally hard for AI.

1. Easier: a high five or a fist bump. The palms or fists meet flat, with few interlaced fingers.
2. Medium: a handshake. The hands grip each other, but the fingers keep a recognisable order.
3. Hardest: interlaced fingers or holding hands with fingers laced together. This is where the depth order collapses fastest.

If you do want a tricky grip, place the camera slightly to the side rather than straight on. From a light angle the fingers sit less directly behind each other, so the model keeps the layers apart more easily.

Fix the hands afterwards instead of rerolling

If the rest of your image is good but only the hands are off, repair that area specifically instead of rerendering the whole photo. Mask only the hands in the photo editor and let the AI fill that patch again, with a short description of the correct grip. The rest of your composition stays intact.

Two things that help up front:

• Work from a reference. Feed a photo of the handshake or high five you want as a start image through image-to-image in the photo generator. A real structure as an anchor beats words alone.
• Test on a cheaper model. Contact poses rarely land on the first try. Run a few cheap variations to find the pose, then finish only the winner on a premium model. You pay per render, so trying costs little.

Frequently asked questions

Why does a single hand work but two together fail?

With one hand the model only has to get that single shape right. With two hands, occlusion is added: at the contact point fingers vanish behind each other and the depth order is lost, so they merge.

Which contact pose is easiest for AI?

A high five or a fist bump, because the palms or fists meet flat with few interlaced fingers. A handshake is medium, interlaced fingers are the hardest.

Do I have to remake the whole photo if the hands go wrong?

No. If only the hands are wrong, mask that area and regenerate it with inpainting. That keeps the rest of the image and saves renders.

Ready to make a believable contact pose? Create an account and test your handshake or high five on a cheaper model before you finish it.