To check that an AI jewellery model generator shows a ring and a chain at believable size, photograph your own piece beside a known reference (a ring of a known size, a chain of a known length), generate the model image, then measure ratios on screen against the real item. Reject any image that drifts more than about 10 to 15 per cent.
The generators do not know the size of what you give them. They redraw it to suit a hand or neck they have invented, so the test must come from you. The method below uses a ruler, a standard card and a spreadsheet, and it works with whichever tool you use.
What goes wrong with scale, in plain terms
Image generators create pictures that look plausible, not pictures that are measured. When you upload a ring and ask for it on a model's hand, the tool decides how large the hand is in the frame, how wide the band should look and how big the stone feels. It has no idea your band is 2 mm wide or your solitaire is 5 mm across.
Three drifts show up again and again. Bands thicken, because a wider band reads better in a small image. Stones inflate, because a bigger stone looks better. Chains lengthen or shorten, because the neckline and pose the tool chose decide where the chain falls, not your clasp-to-clasp length. None of these is a bug in one product; they come from how the pictures are made.
One tool page I checked (SellerPic, October 2026) lists a virtual try-on for jewellery with separate tools for necklaces, earrings, rings, bracelets and watches, and a free plan of 20 credits a month. Its page does not say how it handles real-world size, and one testimonial it shows notes that a perfect result takes some patience. That is a fair summary of the whole category: useful, and in need of your measurement.
If you are still choosing tools, AI tools for a jewellery business, compared by the job they do sorts them by task. This tutorial assumes you have one in front of you and want to know whether to trust its output.
Build a test pack before you open the tool
A test pack is a set of items with sizes you know for certain. Make it once and reuse it for every tool and every update.
- A plain ring of known size. A band you can measure: inside diameter 17.3 mm is the standard measurement for a size 7 on the numeric scale, and a 2 mm wide band makes the sums easy. Measure your own ring with callipers or a ring sizer and write down the inside diameter and the band width.
- A stone ring. A ring with a round stone of known diameter, say 5 mm. This tests stone inflation.
- A chain of known length. A chain you can lay flat and measure, for example 18 inches (45.7 cm), with a pendant of known height, say 20 mm.
- A standard card. A bank-style card measures 85.6 mm by 54 mm, a reference nearly everyone can lay beside a piece.
- A real worn reference photo. Photograph the ring on a hand and the chain on a neck, in daylight, square on, with the card or a ruler in the frame. This is the truth you compare the AI image against.
Spend twenty minutes on this. It is the part nobody does, and it is why most scale errors reach the shop.
The ring test: three measurements
Use the ring with the 2 mm band and 17.3 mm inside diameter, worn on a hand in your reference photo. Then generate an AI image of the same ring on a model's hand, face-on to the camera if the tool allows it.
Measurement 1: band width to ring diameter
Open both images in any viewer that shows pixel positions, or print them and use a ruler. Measure the visible band width and the outside diameter of the ring in each image. On the real ring, outside diameter is the inside diameter plus twice the band thickness (17.3 mm plus a 1.5 mm thickness on each side gives about 20.3 mm), so band width as a share of outside diameter is about 10 per cent in the 2 mm example. If the AI ring shows a band that is 14 or 15 per cent, it has thickened the band by about half again.
Measurement 2: stone to ring
Divide stone diameter by ring outside diameter. With a 5 mm stone on a 20.3 mm ring the ratio is about 0.25. An AI image where the stone is nearly a third of the ring has inflated it.
Measurement 3: ring to finger
Compare ring width, edge to edge, with the width of the finger it sits on. In your reference photo, say the ring is about 1.1 times the finger width. If the AI hand makes the ring look 1.4 times the finger, it is oversized or the finger is slimmed. This one catches the "dainty hand" problem, where the hand is made unrealistically narrow and the ring looks substantial.
A spreadsheet helps. Make columns for the three ratios, one row for your reference photo and one row per generated image, and a final column that shows the percentage difference. Colour anything over 15 per cent red.
The chain test: length, drop and pendant
Chains are harder than rings because the length you see depends on neckline, pose and lens. Rely on three checks, not one.
Check 1: where the chain falls
Retail conventions describe where common lengths tend to sit on an adult: a 16 inch chain around the base of the neck, an 18 inch chain at or just below the collarbone, a 20 inch chain a little below that, and 24 inches towards the breastbone. These vary with body and neckline, so use them as a rough guide and compare against your own worn photo. If your 18 inch chain in the reference photo rests on the collarbone and the AI image drops it to the chest, the tool has lengthened it.
Check 2: pendant to chain width
A pendant 20 mm tall on a chain with links 1.5 mm wide is about thirteen link-widths tall. Measure the pendant height divided by the chain width in both images. If the AI pendant is thirty link-widths tall, it has become a statement piece.
Check 3: pendant to neck
The pendant height against the width of the neck at the collarbone gives a ratio you can compare with the real photo. Allow some wobble for pose, but a pendant that appears twice as large relative to the neck is wrong.
One band and one chain measured across three AI images
Take an illustrative small shop selling a plain 2 mm band in a size 7 and an 18 inch chain with a 20 mm pendant. The owner makes reference photos, then generates three AI images of each piece in the same tool.
| Measure | Real reference | AI image 1 | AI image 2 | AI image 3 |
|---|---|---|---|---|
| Band width / ring diameter | 0.10 | 0.11 | 0.15 | 0.09 |
| Ring width / finger width | 1.10 | 1.15 | 1.45 | 1.05 |
| Pendant height / link width | 13 | 14 | 22 | 13 |
| Chain rests at | Collarbone | Collarbone | Mid chest | Collarbone |
Image 1 is within 10 per cent on every line and passes. Image 3 passes too. Image 2 fails on all four lines: the band is thick, the ring looks oversized, the pendant is nearly twice as tall as it should be, and the chain falls on the chest. The owner publishes images 1 and 3 and discards image 2 without needing to argue about whether it looks nice, because the numbers say it is a different piece.
Notice what made this quick: three ratios per image, a threshold agreed in advance, and a simple pass or fail. Four minutes per image is realistic once the sheet exists.
Feed the tool a better input photo
Half of scale drift starts with the photo you upload. A ring photographed at a steep angle with a wide lens looks thicker at the front and thinner behind, and the generator copies that distortion onto the model's hand. Three habits improve the input.
- Shoot square on. Photograph the ring face-on and a second shot side-on. If the tool accepts only one image, use face-on for rings and flat-lay for chains.
- Keep the lens long. A phone's standard camera at arm's length distorts close objects. Step back and crop, or use the 2x lens, so the band keeps its true proportions.
- Put a ruler in the upload. Some tools ignore it, but a few will keep it in the scene or at least keep the piece beside something of fixed size. Remove it afterwards if it appears in the result.
Before and after: one ring, two inputs
Say a shop uploads a 2 mm band shot at a steep angle with the phone four inches from the ring. The generated hand shows a band that measures 0.16 of the ring's outside diameter against a real 0.10, so it fails. The owner reshoots square on from half a metre away, crops in, and uploads again. The second image measures 0.11 and passes. Nothing about the tool changed; only the input did. When a result fails, reshoot before you blame the generator.
Prompts that reduce drift, and what they do
You cannot make a generator accurate through wording alone, but you can reduce drift. Here are two prompt patterns with what to expect from each. The outputs described are illustrative.
Weak prompt:
"Elegant woman wearing a gold ring, close-up, soft light."
Better prompt:
"Close-up of a hand, relaxed fingers, wearing a slim 2 mm plain gold band
on the ring finger. Band must be thin, about one tenth of the ring's
outside diameter. No other jewellery. Natural hand proportions,
neutral background, daylight."The weak prompt usually gives a chunkier, more decorative ring, because "elegant" invites ornament. The better one tends to keep the band thin but may still thicken it by a few per cent. That is why you measure anyway. Write the ratio into the prompt only as a request, not a guarantee.
For chains, name the length in words and in position: "an 18 inch chain resting at the collarbone, with a small 20 mm pendant about as tall as the width of two fingertips". Again, the tool treats this as a suggestion. Measure the result.
Three more checks that catch quiet errors
The mirror test
Flip the generated image horizontally. If the ring or pendant looks different when flipped (an engraving that reads backwards, a clasp on the wrong side), the tool invented details. This does not measure scale, but it shows how much the tool has redrawn.
The two-ring test
Generate an image with a thin ring and a thick ring on the same hand. If the tool makes them the same width, it is not respecting your product at all.
The earring and bracelet extension
Earrings and bracelets need their own references. For a stud, use the real diameter against the ear in a reference photo. For a bracelet, use the wrist width against the bangle's internal diameter. The same three-ratio approach applies.
A quick sum for bracelets
A bangle with an inside diameter of 60 mm has an inside circumference of about 188 mm (60 multiplied by 3.14). A slim adult wrist often needs a bangle that slips over the widest part of the hand, so the image should show the bangle moving with a little room, not clamped tight to the wrist bone. If the generated bangle looks snug in the picture but your product listing says it slips on easily, one of them is wrong. Compare bangle diameter to the wrist width in the reference photo, using the same ratio method.
For a pair of drop earrings 30 mm long, the earring length divided by the ear's height gives a ratio you can compare. In practice, pick the ear lobe width as your yardstick, since it stays visible whatever the hairstyle.
What failure looks like in the wild
These illustrative cases show how a scale error reaches a customer and what it costs.
- The signet that grew. An AI hand with narrow fingers makes a small signet ring look chunky. Customers order it expecting a statement piece and return it as too small in presence. Cost: shipping both ways and a lost sale.
- The choker that became a necklace. A 16 inch chain generated on a model with a low neckline appears to hang at the chest. A buyer who wanted a collarbone length orders the 16 inch and finds it tighter than expected.
- The stud that ballooned. A 4 mm stud pictured on a close-cropped ear looks like a 6 mm. The listing says 4 mm, the picture says 6, and the customer believes the picture.
- The missing clasp. The generator tidies away the clasp and the chain looks unjoined. The shop's description mentions a lobster clasp, and the mismatch triggers a complaint.
Each case could have been caught with one measured ratio and one real photograph.
Setting your own tolerance
Fifteen per cent is a starting point, not a law. A good rule is to tighten the tolerance as the price of the piece rises and as the buyer's need for exact size grows. For a fashion ring at a low price, 15 per cent on the band ratio is tolerable. For an engagement ring, where size and stone presence are the whole purchase, use 5 per cent or skip AI model images for that line altogether and show the real piece on a real hand. Write the tolerance for each product type on your scorecard so that whoever approves images next month applies the same standard.
Scorecard to keep beside the tool
Piece: ______________ Tool: ______________ Date: ________
Reference photo taken? [ ] yes
Reference ratios recorded? band/ring ___ ring/finger ___ pendant/link ___
Generated image number: ___
Ratios within 15 per cent? [ ] pass [ ] fail
Chain rests where the real one does? [ ] pass [ ] fail
Details match (clasp, engraving, stone cut)? [ ] pass [ ] fail
Decision: [ ] publish [ ] regenerate [ ] use real photoKeep the sheets. If a customer complains, you can show what you checked, and over time you will see which tool and which prompt stay close to the truth.
Where to place AI images and where to keep real photos
Even a passing AI image should not be your only picture. A sensible listing order for a ring is: the real product photo on a plain background first, a real photo beside a ruler or card second, then AI lifestyle images after. For colour and metal accuracy, jewellery photography with AI backgrounds covers how to keep the piece itself true to life, and what AI gets wrong in descriptions of metals and stones checks the words beside the image.
Whether and how to tell shoppers that an image is AI-made is a separate question, covered in the disclosure tutorial for jewellery sellers. Do the scale test first, then decide on the label.
When to stop using the generator
If a tool fails the test on three attempts in a row for the same piece, switch method for that piece: use a real photo on a model friend or hire a photographer for the key lines. For the cost side, AI product photos against a photographer helps you work out whether the generator is saving money once rejected images are counted.
Count your rejection rate honestly. If one in three images fails, the cost of the generator is the cost of three images for every one you publish, plus your measuring time. That is still cheap compared with a photoshoot for some businesses, and not cheap enough for others.
If you would rather build this check into your whole listing routine with someone, that is what my AI implementation consultation covers; you can get in touch here.
Scale questions jewellers ask about AI model images
Why do AI images make rings look bigger or smaller than they are?
The generator redraws your piece to fit a hand it invented, and it has no measurement for the real item. It tends to favour a flattering, balanced look, so thin bands thicken, stones grow and chains lengthen. The fix is to test against a known ring size and chain length and reject images that drift.
Is a small scale error really a problem?
Yes, if it changes what a buyer expects. A ring that looks one size thicker on screen than in the box is a returns risk, and an image that makes a 16 inch chain look like a 24 inch one is a misdescription. Keep the tolerance tight and write it down.
Can I add a size line to the listing to cover myself?
A size line helps, but it does not rescue a misleading picture. Use real measurements in the listing and keep at least one unedited photograph of the piece beside a ruler or a standard card, so the true size is visible.
How often should I rerun the test?
Run it once per tool and again whenever the tool changes its model or you change the prompt. Generators update often, so a result that passed in the spring may drift later. Keep your test pack and last passing images in one folder.
Further reads
- AI for Jewellery Design: Concept Sketches to Client-Approved Renders — Concept sketches to client approval for custom pieces.
- ChatGPT Prompts for Jewellery Shops: 8 for Listings and Clients — Prompts for listings and client messages.
- AI for a Jewellery Business: Anniversary and Repair Client Messages — Anniversary and repair messages for existing customers.
- AI Tools for Ecommerce Photography: Backgrounds, Shadows, Consistent Sets — Backgrounds and shadows for a consistent product set.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: SellerPic homepage (free plan of 20 credits a month, checked October 2026); ring size and card dimensions from standard published conventions. Other tools were not tested here, so run the test on whichever you use.