A supplement bottle carries more text than most web pages: an ingredient list, a dosage line, a warning, a batch code, a barcode, a brand mark. Every one of them is small, and small is what an image model handles worst. It has seen a great deal of text and knows what text looks like from a distance; asked to keep a label while it redraws a scene, it produces something text-shaped — the right length, the right weight, wrong letters. The customer who zooms in finds a word that does not exist.
This page is a method for products with fine text and labels in AI product photography. It publishes no legibility figure for any tool. What it gives is a way of working in which the label is never redrawn — its pixels come from a photograph and stay — and a check that reads every character rather than judging whether the label "looks right".
How do you keep label text legible in AI product photography?
Photograph the label as if it were a document: straight on, filling the frame, evenly lit, in focus, at full resolution, as well as the product as a whole. Use workflows that keep the product's own pixels — a background swap, a studio ground, a scene placed around the product — and never a re-render that draws the product from a description. Generate at the top resolution and compare the label to its photograph at full size, reading it: every word, every number, the barcode's bars, the logo's shapes. A label that is almost right is a rejected image; regenerate from the master rather than retouching the text.
What counts as fine text, for this purpose?
Any text or mark on a product small enough that a customer would zoom to read it: ingredient and nutrition panels, dosage and warning lines, batch and expiry codes, barcodes, care labels, size tags, serial numbers, small logos and wordmarks, embossed or printed type on a cap or a clasp. It differs from a large print or a pattern in being read rather than seen, which means the check is reading, and a plausible approximation fails it.
Why text fails
- Small areas are simplified first. Type is a small share of the pixels; when a product is redrawn, small detail is what gets smoothed.
- The model knows what text looks like, not what it says. It reproduces the rhythm of a line — length, weight, spacing — and invents the letters.
- Numbers drift. A dosage, a weight, a percentage on a label is a number the model has no reason to keep; it becomes a different plausible number.
- Barcodes become stripes. The pattern reads as decoration and is regenerated as decoration.
- Curved labels foreshorten. Text wrapping a bottle is compressed at the edges; a redraw straightens or garbles it.
- Embossing has no colour. Type raised in the same material shows only as light and shadow, and a relit scene changes it.
Shoot the label as a document
- Straight on, filling the frame. A second photograph of each label, the camera square to it, the label as large as the frame allows.
- Even, soft light, no glare. Glossy labels reflect the light source as a white bar across the text; a large soft source, angled away, avoids it.
- In focus at full resolution. The type is only as sharp as the pixels; upload at the camera's original size — up to 10 MB in Photta.
- The whole product too. The document photograph is for the inspection; the master of the whole product, on a plain ground with a colour reference, is what the images are made from.
- Curved labels from the front and both sides. So every line has one photograph where it is not foreshortened.
- Embossed type raked. Light from one side, low, so the relief shows.
Use workflows that keep the pixels
- Background swap and studio ground. A background remover cuts the product out with its label intact and places it; a studio workflow puts it on white or grey with a shadow. The label's pixels are the photograph's. In Photta: the background remover and studio product photography.
- Scenes placed around the product. A lifestyle scene generated around a kept product is fine; the label survives because it is never redrawn.
- Never a one-step re-render. "This bottle on a marble counter" as a single description draws a new bottle with a new label. It is a mood board, not a product image.
- Top resolution for the inspection copy. Read the label on the largest output; a 4K copy is where a batch code is legible.
- Upscale only from the kept pixels. An upscaler sharpening real type is fine; one asked to invent detail on a redrawn label is not.
Document, keep, read
The loop the checklist below serves, named so the other hard-product pages can point at it.
- Document — every label photographed straight, sharp and filled, beside the master of the whole product.
- Keep — workflows that carry the product's pixels into the image, never a re-render.
- Read — every character of every label compared to its document photograph at full size, before publishing.
Fine text and label checklist
| Check at full size | Pass | Fail — regenerate |
|---|---|---|
| Words | Every word as printed, letter for letter | Text-shaped marks; a word changed; a letter swapped |
| Numbers | Dosage, weight, volume, dates, codes identical | A different plausible number |
| Barcode | The same bars and spacing; the number beneath identical | Regularised stripes; a bar count that changed |
| Logo and wordmark | The mark's own shapes and letterforms | Rounded, simplified, re-spaced |
| Curved text | Follows the product's curve as photographed | Straightened, stretched or garbled at the edges |
| Embossed type | Relief readable under the scene's light | Flattened, or reading as printed |
| Label edges and colour | The label's outline and paper or film colour unchanged | A label resized, re-coloured or moved |
| Legibility at listing size | Readable where the channel shows it at zoom | Readable only in the master |
| Method | Product pixels kept; scene added around them | Re-rendered from a description |
A label that fails any row fails the image; a customer reads labels, and a wrong dosage line is not a cosmetic error.
One supplement bottle, one listing
A supplements brand photographs a bottle on white with a grey card, front and three-quarter, and then each label straight on: the front panel, the ingredient panel, the barcode side. The main image comes from the studio workflow — the bottle cut and placed on white with one shadow — and the inspection copy at 4K is read against the document photographs: the brand mark, the product name, the dosage line, the count, the batch code, the barcode's bars. The gallery adds a pale-stone daylight scene generated around the kept bottle; the ingredient panel is read again, because a scene changes light and a relit label can lose contrast. The one failure is a barcode whose bars thinned in the scene; that image is regenerated from the master, not retouched, and read again before it goes live.
Frequently asked questions
Does AI product photography keep label text readable?
It does when the label's pixels come from a photograph and are never redrawn: a background swap, a studio ground or a scene placed around the kept product leaves the label as it was shot. It does not when the product is re-rendered from a description, which draws text-shaped marks. The check is reading the label against its photograph, not judging its look.
Why does AI change numbers and words on my labels?
Because small type is a small share of the pixels and is simplified first, and because a model reproduces the rhythm of a line rather than its content. Numbers have nothing anchoring them and drift to other plausible numbers. Keeping the product's pixels prevents it; reading every character catches what remains.
How should I photograph a product with a lot of small text?
Twice: the whole product on a plain ground with a colour reference, and each label straight on, filling the frame, evenly lit without glare, in focus, at full resolution. Curved labels from the front and both sides; embossed type with low light from one side.
Can I fix a wrong label in an editor instead of regenerating?
No. A retouched dosage line or barcode is a fabrication on the part of the product a customer relies on, and it sits under a scene's light it does not share. Regenerate from the master with a workflow that keeps the product's pixels, and read the label again.
What resolution do I need for legible labels?
The camera's full resolution on the way in, and the top output resolution for the copy you inspect and for any channel that offers zoom. A label legible only in the master is not legible on the listing; check it at the size the channel shows.
What this page does not claim
That any tool preserves text at a given rate, or that legible labels change returns by any amount; this page measures neither. What it claims is that a label is read rather than seen, that its pixels have to come from a photograph, and that the only check that means anything is reading every character.
To keep a labelled product's pixels through a background change, see the background remover; for the general check behind every kept-pixel workflow, how to replace product backgrounds without changing the product; for marks on garments rather than packaging, how to preserve logos and patterns in virtual try-on.