The hand deforms as soon as it holds the product

The extra finger is not what kills a product shot. The failure that actually ships is the bottle that dents where the thumb lands, or the label that stretches around a knuckle that is not quite a knuckle. The rest of the frame can be excellent. The moment skin meets a rigid branded object, both sides of the contact start to average into each other.

An empty hand has one object to get wrong. A grip has two, they occlude each other, and one of them (the SKU) has a geometry you cannot afford to invent. Re-rolling trades the lighting you already like for another lottery ticket on the contact. Fix the hold. Leave the picture.

Why contact is harder than a bare hand

Image models pattern-match hand appearance. They do not carry an anatomical prior, or a CAD model of your bottle. Hands occupy few pixels, appear in enormous pose variety, and are often occluded in training data, which is why extra fingers still show up on stronger 2026 models. Contact adds three constraints the sampler is not jointly solving:

The product is informational, the hand is textural. A cheek or a sleeve can be reconstructed as plausible texture. A label, a cap thread, a carton corner cannot. The moment the model is allowed to re-draw the object in order to make the fingers look attached, it will "attach" them by melting the object.

Occlusion at the contact line is the worst case. Fingers wrap, hide each other, and hide the product. From the camera's point of view the contact is a tangle of edges. Averaging across that tangle produces a blob that is approximately a hold.

The two objects have different correctness tests. A hand can be plausible. A SKU is either the SKU or it is a different object. Partial credit does not exist on the product side, which is why a grip that looks fine at thumbnail size fails the moment someone reads the label.

The general hand-repair loop (crop in, mask, describe the pose, iterate on denoising strength) is covered in fixing hands without regenerating the image. What follows is the grip-specific version: protect the product, repair only the contact, and fall back to a composite when the object itself starts to move.

Mask the hold, not the product

The instinct is to circle the whole hand-plus-bottle cluster and ask for "a natural grip." That hands the model permission to rebuild the SKU. Draw the mask so the product surface stays outside it wherever the surface is visible.

1. Crop to the hold plus a band of context. Wrist, cuff, the product, and a little of the surface it sits on. Editing at that scale is what gives the region a real share of the model's budget. A grip that was two percent of a lifestyle frame is now most of the canvas.
2. Mask the fingers, the visible palm, and the contact shadow. Include 20 to 30 pixels of wrist or cuff so the new hand has somewhere to join. Do not include the product body, the label, the cap, or any printed mark you can still see.
3. Go asymmetric at the contact line. Generous and feathered on the flesh side. Tight, almost hard, on the product side. A halo lives in a sliver; a warped label lives in a sliver you gave away. This is the opposite of a removal mask, where you want the model to rebuild the surrounding surface.
4. Prompt the geometry of the hold, not the noun "hand." Name which hand, how many fingers are visible, where each one lands, and that the object is rigid and unchanged.

A prompt that usually lands:

Right hand gripping a matte black 200ml glass dropper bottle. Thumb on the front of the bottle below the label, index and middle fingers wrapped behind, ring and little finger folded against the palm. Five fingers, distinct knuckles, skin compressing slightly against the glass. The bottle shape, label and dropper cap stay exactly as they are. Soft window light from camera left.

What that prompt is doing: a count, named contact points, a camera-side view of the hand, and an explicit instruction that the object is not in play. "Natural hand holding the product" is the weak version. It re-opens the same pose space that failed the first time, and it does not tell the model the product is off-limits.

Set denoising strength in the middle of the range (around 0.5 to 0.7 on a 0–1 scale). Too low and the fused contact survives under a slightly different texture. Too high and you get a new hand and a new bottle. Keep the seed fixed while you iterate on wording; changing both at once means you cannot tell which one moved the result.

Two or three tries is enough. If the third still warps the SKU, stop. The model is using the product as a blending surface. Go to the composite.

Drive the same loop from the photo editor. Image-to-image takes a mask alongside the prompt and the seed, so unmasked product pixels stay as they were. Element re-render versus masked inpainting is the decision behind that choice. Naming "the hand" is ambiguous the moment a second hand, a sleeve, or the product itself could be read as part of the same object. A mask is the hard boundary.

When the product still moves, composite the real SKU

Inpainting the grip assumes the generated product is close enough to keep. When it is not (the bottle has gained a waist, the carton has rounded off, the label copy has started to drift) you stop asking the model to respect geometry it never had, and you supply the geometry.

The fallback is a three-layer composite, not another generation:

1. Keep the generated frame as the plate. Lighting, set, wardrobe, everything that was already right.
2. Drop a real photograph of the SKU over the deformed object. Match scale, rotation and perspective to the hold. A three-quarter product shot from your packshot set is the usual source; a straight-on hero is harder to seat in a lifestyle angle. Neutral light on the photo is easier to grade into the plate than a photo with its own hard studio key.
3. Inpaint only the fingers so they meet the real silhouette. The mask now has a correct edge to wrap around. Prompt the contact against this object: "thumb pad flattening on the front of the glass, index finger visible behind the bottle, no change to the bottle."

A few minutes on perspective and a colour-match is cheaper than a fourth full generation, and it is the only version of this job that can guarantee the label still says what the physical product says. For anything you will reuse, treat that photograph as a reference image as well as a composite layer. A deformed grip that gets copied into a reference set will be reproduced in every downstream frame. Holding a physical SKU steady across generated shots is the same discipline applied to video.

A phone photo of your own hand, at the same angle, in roughly the same light, is the other reference worth making. One hand, plain background. Pass it with the masked edit so the model is no longer inventing a pose from words.

Prevent the contact failure before you generate

Repair is the right move on a frame you already like. It is still more expensive than not creating the failure.

• Give the hand a rigid object, then protect that object. A hold is easier than an empty gesture only if the object is allowed to stay itself. Do not ask the sampler to invent the SKU and the grip in one pass if you have a packshot you can composite.
• Do not make a wrap-around close-up the hero. Fingers fully encircling a small bottle, seen edge-on, with the label facing camera, is close to the worst-case pose. A three-quarter hold with two fingers and a thumb visible, label unobstructed, is a smaller ask.
• Keep the label sharp and let the fingers fall off if the shot allows it. Shallow depth of field over the knuckles is optics, not a cover-up.
• Check the contact first when you review. Look at where skin meets the object before you evaluate the rest of the frame.

If the brief does not require a grip, do not stage one. Product on the table, open palm, or a packshot with a hand in the background out of focus all remove the contact problem rather than repairing it.

FAQ

Why does the product dent even when the fingers look fine?

Because the sampler is solving one region, and the cheapest way to make a hold look attached is to pull the object toward the hand. The fingers can land on a count of five and still "join" by warping the bottle. That is why the mask has to stop at the product surface. If the object is inside the mask, you invited the dent.

Should I mask the whole product and regenerate the grip as one cluster?

No. That is how the label copy drifts and the cap changes shape. Mask flesh and contact shadow. Leave every visible product pixel unmasked. If the generated SKU is already the wrong object, do not inpaint a bigger region to "fix" it. Composite the real product and inpaint only the fingers against that silhouette.

Can I skip the mask and just say "do not change the bottle"?

Sometimes, when the product is large, high-contrast, and unambiguous in frame. It fails as soon as the bottle shares colour with the sleeve, or there are two similar objects, or the hand occludes half the label. A sentence is not a hard boundary. Inpainting is.

When is it faster to re-stage the shot without a grip?

When the hold is not the brief, or when one composite attempt still warps the SKU. The grip is not mandatory just because the first prompt mentioned one.