Hands fail because hands are the hardest thing in the frame.

A hand is a small region of the image with an enormous amount of structure: five articulated chains, a palm with its own geometry, and usually another hand or an object interacting with it.

It is also almost always occluded. Fingers overlap each other, the palm hides the lower joints, and the hand disappears into a sleeve or behind an object, which means the parts that would explain the pose are not visible.

Training data makes this worse rather than better. The web is full of hands in the same few poses, and far less of the unusual angles that photographs and illustration actually need.

So the failure is not mysterious or arbitrary. It is a small, complex, self-occluding object being reconstructed from ambiguous evidence, which is also why it survives longer than faces do.

The errors repeat across models and styles. Knowing the taxonomy is most of the inspection.

Start at the thumbnail, then zoom. Errors that are obvious at full size can vanish at publication size, and an image that reads perfectly small does not need to be regenerated for a landing page.

Zoom to the hands and follow a single finger from tip to wrist. The task is not to judge whether it looks right, it is to trace a continuous, plausible structure.

Then look at the join between hand and object. Grip is where generators fail most consistently, because the object and the hand were never modelled as touching, only as overlapping.

Finally compare the hands with the rest of the image: light direction, shadow softness, lens blur and colour. An anatomically plausible hand with the wrong lighting is still a compositing error.

Handshakes and clasped hands, because two hands interlock and the generator must decide which fingers belong to whom.

Playing an instrument, because every finger has a required position that a viewer may know.

Holding small objects between finger and thumb, where the pinch is a precise geometry nobody notices until it is wrong.

Hands near the face, where they interact with hair, glasses and the jawline, and where the viewer's attention already is.

Poses from the back or in profile, where the palm is invisible and the finger count must be inferred.

If a prompt requires one of these, budget for several attempts or plan to composite a real hand.

First, regenerate with the hands described explicitly: both hands visible, fingers relaxed, holding nothing. Removing the difficult interaction solves more cases than any inpainting.

Second, inpaint only the hand region with a prompt that describes the hand and its task, keeping the rest of the image fixed. This works when the pose is simple and fails when it is not.

Third, crop the hands out. A portrait ending at the elbow is a legitimate image, and it is cheaper than a repair that leaves a visible seam.

Fourth, composite a photographed hand. This is the reliable option for commercial work, and it takes a retoucher twenty minutes with a decent plate.

Most hand failures appear because the brief demanded something hard without knowing it. Hands in motion, hands with props, hands overlapping other hands, hands at odd angles.

Describe the pose in plain physical terms instead of a mood: fingers curled around a mug handle, thumb on the near side. Specific mechanical language reduces the guessing.

Or remove the hands. Pockets, a table edge, sleeves, a crop at the wrist: each removes the hardest region and often improves the composition, since hands in generated images are frequently the least considered part of the frame.

For a set of images across one brand, standardise the poses that work and reuse them. Consistency is worth more than variety when the alternative is a redraw per asset.

At 200 pixels wide in a card, nobody counts fingers. If the image is decorative, cropped small and never inspected, a six-fingered hand is not a defect worth a regeneration.

The threshold is consequence. A hero image, a print asset, a medical or instructional illustration, anything a client will look at closely: these need the hand pass.

Ask what the viewer is looking at. If the hands are the subject, or beside the subject, errors are noticed. If they are in the background of a wide shot, they are not.

Write the threshold down for your own assets. Ad hoc judgement means the same image passes one week and fails the next, which is a process problem dressed as a taste problem.

A visible hand error is strong evidence an image was generated, which makes it useful when you need to know. It is not proof of the reverse: a clean pair of hands tells you only that the region is plausible.

Editors use it as a first-pass filter because it is fast and requires no tooling. Step one is the hands, step two is text, step three is the geometry of the scene.

It also cuts the other way. A retouched photograph can show odd hands because of cropping, montage or a poor composite, and accusing a photographer of using a generator on that basis is a mistake with a real cost.

Treat the hand check as a reason to look closer, which is what any single artefact deserves.

Build the hand check into the review step, at the resolution the asset will ship in, with one person responsible for signing it off.

For volume, a contact sheet at full-crop of every hand region is faster than opening each image. Ten crops on one page, checked in a minute.

Record the outcome per asset, because the same failure returns when the same prompt is reused. A note saying remove hand prop is worth more than a general instruction to check carefully.

And keep the reusable plates. One photographed hand on a neutral background solves a surprising number of future briefs.

Count the digits. Check the thumb. Trace one finger to the wrist. Follow the outline of whatever is being held. Compare the lighting on the hand with the lighting on the face.

Five checks, at full zoom, before the image goes anywhere. It catches most of what a viewer will find later, and it is fast enough that it actually gets done.

Use the inspection list on this page to run it consistently, and to record what you found so the same brief does not fail the same way twice.

The hands are not the whole of image quality, but they are the cheapest place to start looking.