AI product photos get accessory size wrong because the reference photo carries no scale information: a 12 cm clutch and a 45 cm tote on the same white background are geometrically identical, so the model renders the size its training data considers typical. The fix is stating size in terms the model can use — the human body. “45 cm chain” means nothing to an image model; “sits just below the collarbone” is a rule it can follow and a reviewer can check. CatalogX automates that translation: type the measurement, get the body anchor.
The scale problem, precisely
Put a mini clutch and a weekender bag on the same white sweep, crop them to the same frame, and the two photographs contain the same information. Nothing in either image says which one fits in a palm. Garments partially escape this — a kurti’s outline implies a body — but a bag, a chain, a watch face or a pair of sunglasses implies nothing.
Image models fill that gap with the statistically typical answer. Typical is exactly what a seller of an unusually large tote or an unusually delicate chain does not want, and the buyer’s disappointment arrives as a return: for bags and jewellery, the size-and-material surprise is the dominant return reason cited across e-commerce photography guides.
The research measures the same thing. The GenScale benchmark (arXiv 2609.00525, 2026) evaluates relative object scale in image generation and editing, and the practitioner method that survives contact with it is consistent: pair the metric dimension with a size-matched human reference — a hand for objects under 30 cm, a torso for 30–60 cm, a full body above — because bodies are the scale reference image models handle best.
The body is the only ruler in frame
An on-model photo exists to answer two questions a packshot cannot: how big is this, and how does it sit on a body. That only works if a body part of known size is actually in frame, unobstructed, next to the product. A close-up so tight that the bag has nothing of known size beside it has failed at the format’s whole job, however sharp the leather is.
This is why CatalogX anchors every on-model accessory shot to a family-specific landmark by default, whether or not a size was stated:
| Product family | The ruler kept in frame | What “how it sits” means |
|---|---|---|
| Bags | Hand and forearm, torso width | Where it rests, how far the strap drops, how it hangs under its own weight |
| Jewellery | The body part it is worn on, in full | How far the necklace drops below the throat; how far an earring hangs below the lobe |
| Watches | The wrist and hand | How much of the wrist’s width the case covers |
| Sunglasses | The whole face | How much of brow and cheekbone the lenses cover |
| Hats | Head, face and shoulder line | How low the crown sits; how far the brim reaches against the shoulders |
| Belts | The natural waist | Strap width in finger-widths; where it lands on the torso |
| Scarves & dupattas | Shoulders, arms, torso | How far the ends actually fall — the fall is the size |
| Footwear | Feet and lower legs | Heel height against the ankle; depth of the sole |
Necklace lengths: the translation table
Jewellery has the industry’s cleanest size vocabulary, and it is body-anchored already — the standard names describe where the chain sits, not just its length. This is the exact mapping CatalogX applies when a seller types a chain measurement:
| Stated length | Standard name | Where it renders on the model |
|---|---|---|
| Up to 38 cm (≈14″) | Collar | Tight, halfway up the neck |
| 38–42 cm (≈16″) | Choker | Base of the neck, above the collarbone |
| 43–48 cm (≈18″) | Princess | Just below the collarbone |
| 49–61 cm (20–24″) | Matinee | Between collarbone and bust |
| 62–91 cm (28–36″) | Opera | Mid-chest, at or below the bust |
| Above 91 cm (36″+) | Rope | Below the chest |
Earring drops band the same way — a 2 cm drop hangs just below the lobe, 5 cm reaches toward the jawline, 8 cm grazes the shoulder — and the piece kind disambiguates the number: “20 cm” is a collar chain on a necklace and an impossible earring. Watch cases band by wrist coverage (a 36 mm case sits inside the wrist, a 44 mm spans it), hat brims by reach against the shoulder line, heels by the arch they force.
From stated size to checked render
Stating the size fixes the input. The remaining question is whether the render obeyed. CatalogX closes that loop with its accessory reviewer: when a size or measurement was stated, the finished image is graded against it — a bag stated at forearm length that renders palm-sized fails review with the mismatch named, exactly as a kurti stated hip-length fails when it renders at the knee. When the jewellery piece is known, wear location is checked the same way: a necklace rendered as a bracelet is a category failure regardless of how faithful the metalwork is.
Two honest caveats. First, this is expectation-setting, not metrology — the mechanism anchors the render to the right body landmark; it does not promise millimetre precision. Second, a control only helps when it is used: with everything left on auto, size is inferred from the photo, which is the same guess every other tool makes.
What to do with the photo itself
Three photo habits move accessory scale accuracy more than any setting:
- Stuff bags to their carried shape. The render reproduces the silhouette you show — an empty, collapsed bag renders collapsed.
- Shoot one shoe in true side profile. Heel height is measured from your photo; a top-down pair shot has no heel to measure.
- Uncoil chains fully, or state the length. A coiled necklace’s length cannot be read; the measurement box exists for exactly this.
CatalogX’s upload check reads the photo against these rules before the render and warns — a collapsed bag, a missing side profile, a coiled chain, or a photo that does not match the selected product type each get a specific message — so the first credit is not spent discovering them.
Frequently asked questions
The reference photo carries no scale, so the model rendered the typical size for the category. State the real size — either a tier (“mini” through “oversized”, defined per product family) or the exact measurement (“30 × 22 cm”) — and the render is anchored to the body landmark that measurement implies.
State the real chain length and let the mapping do the work: 38–42 cm renders at the base of the neck (choker), 43–48 cm just below the collarbone (princess), 49–61 cm between collarbone and bust (matinee), 62–91 cm at mid-chest (opera).
No — it anchors the render to the right body landmark and the result is reviewed against the stated size, with clear mismatches failing review. It is an expectation-setting mechanism, not metrology.
Use the size tiers, which are worded per product family — a “large” bag is elbow-to-fingertips, a “large” watch spans the wrist, “large” jewellery covers the collarbone. Even a tier beats auto, which is a guess from a photo that carries no scale.
Yes — the same size statement travels through single shots, 20-SKU bulk runs, and every shot of an e-commerce pack, and it is recorded with the shoot so re-runs and edits keep it.