How to get hands right in AI product photos
Brief the grip the way you would brief a hand model: which hand, how many fingers show, where the thumb sits, what the nails look like, and where the frame crops. Give the model a real photo of a hand holding your product, keep the product large and the hand simple, and check every finger. Research traces the errors to how hands bend and hide each other in training photos.
Why AI gets hands wrong
A hand has five digits and many joints, and in photos it bends, overlaps itself and hides behind whatever it holds. The HandRefiner paper (2023) describes the result: image models "suffer from generating accurate human hands, such as incorrect finger counts or irregular shapes," because learning a hand’s structure and pose from training images "involves extensive deformations and occlusions." Model makers still list fine detail as a weak spot; Google says Nano Banana Pro "can still struggle with small faces, accurate spelling, and fine details in images."
Product shots add two problems of their own. The grip has to make physical sense for the product’s weight and shape, and the hand has to be the right size for the product, which the model cannot know from a packshot.
Brief the grip
| Decision | Options | Example wording |
|---|---|---|
| Which hand | Left or right; one hand or two | "her right hand" |
| Grip | Pinch, cradle, wrap, fingertip press, two-hand hold | "cradled in an open palm" |
| Fingers in view | Count them | "thumb and two fingers visible, the rest hidden behind the jar" |
| Thumb | Where it sits and what it does | "thumb resting on the lid, clear of the label" |
| Contact | Where skin meets the product | "fingertips pressing lightly into the soft tube" |
| Nails | Length, shape, finish | "short natural nails, clear polish" |
| Skin and age | Tone, texture, age, jewelry | "a woman in her forties, visible knuckle creases, no rings" |
| Crop | Where the frame cuts | "cropped at the wrist, sleeve out of frame" |
Product: [name], [real height and width in cm], [weight, if it changes the grip] Hand: [left or right], [one or two hands], [age, skin tone, skin texture] Grip: [pinch, cradle, wrap or press]. Fingers visible: [number]. Thumb: [where it sits] Contact: [where fingers touch the product; any pressure] Nails: [length, shape, polish] Crop: [wrist, mid-forearm or fingers only] Label: [stays fully visible, or which part may be covered] Reference: [photo of a real hand holding the real product]
Use a real hand as the reference
The quickest route to a believable grip is a photo of a real person holding the real product. It gives the model the grip, the contact points and the size of the hand against the product in one image. Take it with a phone in soft light, label toward the lens, and send it with the packshot. Say which image carries what: the product from the packshot, the grip from the hand photo.
The same photo does a second job. In Baymard Institute’s testing, a shopper grasped a speaker’s size at once from an image of it held in a hand; see how to show product size in photos. And ask for plausible hands in plain words: the GPT Image 2 prompt behind the trouser store photos we kept in 2026 included the line "Natural believable hands and body proportions."
Crop and compose around the problem
- Crop at the wrist or mid-forearm. A whole arm adds an elbow and a shoulder to get right.
- Show fewer fingers. A cradle, or a wrap seen from behind, hides fingers behind the product.
- Keep fingers apart. Interlaced fingers and two hands touching are where fused and extra fingers tend to appear.
- Choose a resting grip over a hand caught mid-gesture.
- Keep fingers off the label, so the model does not redraw text around them.
- In packshots, keep the hand smaller in the frame than the product.
How to fix a bad hand
- Remake with the same prompt and one change: a simpler grip, or fewer fingers in view.
- Edit only the hand. Many tools can redraw one masked area, but Google warns that masked edits on Nano Banana Pro can produce "unnatural results, visual artifacts, or disjointed scenes," so check the join at the wrist. Research methods such as HandRefiner take the same approach: redraw the hand and leave the rest of the image alone.
- Change the crop so the problem finger leaves the frame.
- Composite a real hand photo taken at the same angle and in the same light.
- Shoot it. For a hero image where the hand carries the idea, a hand model and a photographer can be quicker than a long run of renders.
Hand checklist
- Five digits per hand, counting the thumb, and no extra hands in the frame
- The thumb is on the correct side for a left or right hand
- Every finger bends at its joints, in a direction a finger can bend
- Fingers are separate, not fused together or melted into the product
- Nails match in shape, length and polish, one nail per finger
- Skin tone matches the arm and face, where they show
- The hand is the right size for the product and for the person
- Contact looks real: soft products dent, and fingers cast small shadows
- The grip could hold the product’s real weight
- The label stays readable
- Rings, watches and polish stay the same across the set
Hands are one section of the full sign-off list in the AI image QA checklist.
Questions people also ask
- Should I hire a hand model instead?
- For a hero image where the hand is part of the idea, often yes: a real hand is right by definition. AI makes sense for volume, variations and scenes where the hand is small in the frame. A middle path works well: shoot one real reference hand and let AI build scenes around that grip.
- Which AI model is best for hands?
- The answer changes with each release, and none is reliable enough to skip the finger check. Run the same grip brief on two or three models, ten renders each, and count the keepers. See which AI image model is best for product photos.
- How do I keep the same hands across a set?
- Treat the hands like a face: make one reference of the same person’s hands with the nails, rings and skin fixed, and send it with every shot. See how to keep the same model in every AI photo.
Where Overs fits
For products that are held or worn, Overs makes a character sheet first: one person from several angles, with close-ups, on a plain background, which you approve before any other photo. Later photos with a person use that same model, and you can remake one photo with a note when a hand goes wrong.
Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.