Cosmetic Product Photography With AI: Textures and Honest Colour

Use AI on cosmetic product photos for the things around the product: background, shadow, dust, crop, size and consistent lighting. Leave the colour, finish and texture of the product itself as the camera captured them. Shoot real swatches under controlled light, check the edited file against the physical shade, and log returns with a "colour not as shown" reason.

The temptation is understandable. A one-click enhance makes a muted rose lipstick look richer, and richer sells. It also sells a colour you do not ship. Every shopper who receives a more orange lip or a paler base becomes a return, a bad review or a customer who does not come back. The rest of this tutorial sets limits, a repeatable workflow and a simple numeric check.

Follow me on Instagram@sagnikteaches

Three kinds of AI edit: safe, borderline, and off limits

Sort every edit by whether it touches what the customer is buying. Anything affecting the product's colour, shine, coverage or texture changes the information the buyer uses to decide.

Connect on LinkedInSagnik Bhattacharya

The last two rows have regulatory weight, not just a quality rationale. Published advertising guidance on cosmetics from one national advertising regulator, dated August 2025, says claims, including visual ones, should not misleadingly exaggerate the effect the product can achieve. It treats general minimal airbrushing as likely to be acceptable, but warns against retouching features directly relevant to how the product appears to perform, and says lash inserts in mascara advertising should do no more than replace damaged or missing lashes. It also says a superimposed disclaimer does not stop a misleading advert being misleading. Other markets have their own rules, so treat that guidance as a clear statement of the principle, then check the rules where you sell.

Subscribe on YouTube@sagnikteaches

For general questions on whether AI product images beat a photographer on cost and quality, AI product photos versus a photographer covers it; this tutorial concerns the narrower question of keeping beauty products honest.

Shooting the source image so AI has less to ruin

Most colour error arrives before any software touches the file. A few fixed habits cut it:

1. Fix the light. Use the same daylight-balanced, high-quality continuous lights every time, in the same position, with windows covered. Mixed lighting (a window plus a desk lamp) is the commonest cause of swatches that look different from week to week.
2. Include a colour reference. Put a grey card or a colour chart in the first frame of each session. It lets you fix white balance once and check that skin tones and neutrals stay put.
3. Lock the camera settings. Manual white balance, fixed exposure, same lens and distance. Shoot raw if the camera allows it, and export with one colour profile for the web, usually sRGB.
4. Swatch on real skin, in real range. Photograph each shade on at least three clearly different skin depths, with the undertone noted, and keep one on a neutral card or swatch stick for a no-skin reference.
5. Shoot the texture separately. One macro of the product at rest (a lipstick bullet, a cream in the jar) and one of it applied. Do not use the same photo for both jobs.
6. Record the session. A one-line note: date, lights, white balance, lens, who shot it. When a customer complains of a shade, you want to know what was true on the day.

Where you cannot control the shoot, such as photos from a supplier, treat their colours as unverified and compare with the physical product before you publish. Your brand rules for colour belong in writing; brand guidelines for AI images shows how to set colours, style and limits so every edit follows the same ones.

A worked check: one lipstick shade, three versions of the photo

An illustration. A small brand sells a lipstick called "Rosewood". The bullet photographed under controlled light, with a grey card in frame, gives a swatch whose centre reads RGB (176, 86, 92) when sampled with a free colour-picker tool. The team then makes two edited versions: one where an AI "enhance" is applied to improve the look, and one where AI only replaces the background. Samples are taken from the same spot on the swatch each time.

The distance is the straight-line gap between the two colours, worked out as the square root of (16 squared plus 14 squared plus 4 squared): square root of 468, about 21.6. The background swap is within a hair of the source. The "enhance" has pushed the red up and the green down, producing a hotter, more saturated pink that the lipstick does not have. A buyer expecting the second image receives the first, and "colour not as shown" appears on the return form.

The threshold is yours to choose. I would start with a distance below about 6 on a sampled patch as pass, 6 to 12 as look closely, and over 12 as fail, then adjust it by comparing printed or on-screen results with the real product. That is a working rule, not an industry standard. RGB distance is also a blunt tool: it is a fast screen, not a proper colour measurement, so confirm any shade you are unsure about with a physical swatch beside a calibrated screen.

Making the check routine rather than a one-off

Do not sample by eye. Pick the same two or three patches per product (centre of the swatch, an edge and the shadow side), record RGB for the source and for each edited file in a sheet, and let a formula flag distances over your limit. A person adds the final look at a screen. If you have hundreds of products, ask an AI assistant to write the spreadsheet formula, then test it on three rows you worked out by hand.

Texture: what a photo may claim and what AI cannot supply

Texture is information. A shopper reading "whipped cream texture" looks at the photo to see how thick it is, whether it has shimmer, whether it will sit on skin or sink in. AI texture tools tend to do the opposite of helpful: they smooth, add glossy highlights or invent clean swoops of product that you cannot actually make.

Keep three rules for texture shots:

• Photograph the real product. A balm on a clean swatch card with a spatula lift, a foundation drawn in a line across a card, a cream dollop on a neutral surface. Imperfections are the point.
• Do not retouch the product edge. Cloning out a ragged edge on a cream swoosh changes what the cream does. Remove dust on the backdrop, not defects in the product.
• Label finish in words next to the picture. "Satin finish, buildable coverage" is a statement the picture should support, not replace.

A short, illustrative example. A brand sells a tinted balm and a shopper complains it is not as glossy as the website. The brand's website image was a macro of the bullet with an AI-added highlight. The fix was to reshoot with real light, and add a second photo of the balm on a lip, unretouched, with a line: "Sheer, natural shine, not a high-gloss finish". Returns citing finish on that product fell in the next quarter, though with small numbers you should treat that as a hint, not a result.

Models and skin: where AI edits cross into claims

As soon as skin is in the picture, the product's claim is in the picture too. A foundation advert showing a perfectly even complexion, smoothed by an AI skin filter, invites the question of what the product did and what the software did. The same guidance says filters are not a problem in themselves, but they must not mislead about what the product does, and any "after" image should show what customers can generally achieve. The same logic covers lash lengthening on a mascara model, line removal for an eye cream and a lip plumper shown on retouched lips.

What I would allow for model shots, product by product:

1. Minimal cleanup. Stray hairs, a passing blemish unrelated to the product, a clothing crease. Keep a before and after copy of each file so you can show what was done.
2. No effect-zone retouching. If the product claims to improve lashes, lips, lines or evenness, leave that area as shot.
3. Representative results. Show the product on varied skin, applied the way a customer would, ideally photographed right after application.
4. Keep the originals. Store the unedited image with the edited one for as long as the advert runs. If anyone questions the picture, the originals show what was changed and what was not.

The wording on what you may claim next to the image matters as much as the pixels, and what AI copy must not say about skincare claims is the companion piece. When AI-generated or AI-altered imagery is used in ads or social posts, how to label AI-generated images and video covers disclosure. Marketplaces and ad platforms have their own image policies too, which differ and change, so read them before you upload lifestyle composites.

Where AI genuinely helps, with a prompt you can reuse

Beauty brands with a large range have a boring problem: 120 shades, each needing the same crop, background and margin. This is where AI earns its keep, because the edit is uniform and does not touch the colour. A typical routine:

1. Export every shade as a raw-converted file with the same profile.
2. Run a background cleanup to a plain, light neutral, with a soft shadow under the product.
3. Resize to each channel's format: square for the store, tall for stories, wide for the banner.
4. Run the sample check on the colour patches for every file, flag anything over your limit.
5. A person looks at the flagged ones and a random 10 percent of the rest.

When using a chat assistant with image upload to help with a review, ask for description and comparison, not repair, and never treat its colour judgement as measurement. Do not expect a vision model to tell you reliably that two similar pinks differ by a few units. Use it as a second pair of eyes for composition and artefacts:

I am uploading two product photos of the same lipstick shade: A is
straight from the camera, B is after a background cleanup. Do NOT tell me
which colour is correct. List only visible differences in: edge quality
around the product, any colour tint on the product from the new
background, reflection or highlight changes, texture or surface
smoothing, and anything that looks added or removed. Use plain words
and say "I can't judge exact colour from a screen image" if relevant.

An illustrative reply: "Edge quality: slight halo at the left of the cap. Tint: the bullet looks marginally warmer in B, possibly from the cream background. Highlights: the glossy streak on the bullet is wider in B. Surface: the bullet looks smoother in B." Each of these is a prompt for your own eyes, then for the numeric check. The wider question of what a chat assistant can read from a photo is covered in what ChatGPT can read in PDFs, spreadsheets and photos.

Making the photo carry the shade: names, text and alt text

Colour accuracy does not live only in pixels. Shoppers choose from a name, a number, a swatch dot and a picture at once, and a mismatch between any two sends them to the returns desk. Put the following beside every shade:

• A shade name and number that appear on the pack, in the picture and in the order email, so the match is easy at unpacking.
• An undertone and depth line: "Medium, warm-neutral". For lips: "Muted rose, medium depth, cool undertone".
• A 'looks like on' note from real swatches, such as "Swatched on light, medium and deep skin" with all three images.
• A 'colours vary by screen' note, placed after the image, never in place of an honest one.
• Alt text that names the shade. Shopify's help centre recommends alt text of 125 characters or fewer (the maximum is 512), for example "Rosewood lipstick swatched on medium skin, muted rose with a satin finish". This also helps accessibility and image search.

When writing the product copy to sit beside the photos, make AI work from your swatch notes, not from invention; writing Shopify product descriptions with AI shows the prompt shape.

Shade-mismatch returns: counting what the photos cost

You can test whether honest photos pay, with your own numbers. An illustrative shop sells 1,200 lip and complexion units a month. Before any change, 6 percent come back, and half of those, 3 percent, give the reason "shade not as expected". Each return costs about $9 once you add the return label, the handling time and the unsellable product, though cosmetics often cannot be resold at all and then cost the full product cost too; use your own figure.

• Shade returns per month: 1,200 times 3 percent = 36.
• Cost: 36 times $9 = $324 a month, before lost repeat purchases.
• If a reshoot and a colour check cut shade returns from 3 percent to 2 percent, that is 12 fewer returns, or $108 a month, about $1,300 a year.

A full reshoot of 60 shades with a freelancer might cost more than that in year one. So judge the investment by which shades return most, not by the range average. Work out return rate per shade; if three shades make up half of all colour returns, reshoot those three first. Spotting patterns in complaints, returns and defects with AI covers how to feed the return reasons into an assistant and sort by product.

Capturing the reason at the moment of return

Add a return reason list with "colour not as shown" and, one step further, a free-text line such as "What did you expect?" Ask one follow-up in the confirmation email: "Was it lighter, darker, warmer or cooler than you expected?" Four options give you direction as well as a count. Four answers saying "more orange" on the same shade is a diagnosis: check that shade's photo, its undertone line and its name.

A mistake that looked like good marketing

An illustrative case. A brand adds an AI "light and bright" filter to every foundation swatch photo for a spring launch, because the shots looked flat. Returns on the lighter shades climb, with comments such as "much darker than the picture". The filter lifted exposure across every swatch, so each shade looked lighter on screen than it is in the bottle, and the effect was strongest at the light end, where tones were pushed close to white. Shoppers matched themselves to a picture that was paler than the product they received.

The recovery steps were dull and effective: reshoot with fixed light and the colour reference, drop the filter, add the swatch-on-three-skins images, and track returns by shade for two months. The lesson is not "never edit"; it is that a global edit affects every shade differently, and nobody checks a global edit shade by shade unless the process makes them.

A one-page routine for each new shade or product

1. Shoot under fixed light with the colour reference in frame; note the session settings.
2. Sample the swatch patches in the source file; record the RGB values.
3. Edit with the allowed list only: background, shadow, dust, crop, size.
4. Re-sample the same patches; compute the distance; stop if over your limit.
5. Compare on a real screen next to the physical product, on a phone as well as a laptop.
6. Write the shade info: name, number, undertone, depth, finish, alt text.
7. Log it: original file kept, edit list, who checked, date.
8. Review after 60 days by return reason for that shade, and adjust.

The routine takes about 15 minutes a shade once the lights and colour card are set, which is the time to fix on a busy day. If you want help deciding which photos to reshoot first, and setting the edit rules your team can run on its own, that is what my AI implementation consultation covers.

Questions about AI edits on beauty product photos

Can I use AI to change the background of a lipstick photo?

Yes, a background swap or cleaner shadow is low risk, provided the product itself is untouched. Check afterwards that AI has not shifted the shade of the bullet, tinted it with the new background's colour, or smoothed the surface so a matte finish now reads glossy.

Is it acceptable to generate a swatch image with AI?

Not as the shade a customer will rely on. A generated swatch is an invention, not a record of your product. Photograph real swatches under controlled light and use AI only for cropping, background cleanup and consistent sizing.

How many skin tones should my swatch photos show?

As many as your range needs buyers to judge. For a foundation or tint, show at least three clearly different depths, plus an undertone label, so shoppers can match against their own. For a lip colour, a light, medium and deep arm is a practical minimum.

Does a note saying colours may vary protect me?

It helps customers understand screens differ, but it does not rescue an image that overstates the product. Advertising guidance says a disclaimer cannot fix an inherently misleading image. Fix the image first, then add the note as a second layer.

How do I know if my photos are causing returns?

Add a return reason such as 'colour not as shown' and track it by shade over a month or two. A single shade with several times the average rate points at its photo, its name or its undertone, not at customers.

Further reads

• AI Model Shots for Boutiques: Try-On Images Without a Photoshoot — The same honesty questions for AI model images in fashion.
• AI for Beauty Brands: Product Copy, Photos and Customer Questions — Where photos sit among a beauty brand's other AI jobs.
• How to Automate Returns and Refunds With Clear AI Rules — Set return rules once shade reasons are tracked.
• Is Midjourney Worth It for Small Business Marketing Images? — Judge image tools for ads and lifestyle scenes.
• Taking a Local Shop Online With AI: Product Listings in a Weekend — Get product listings live in a weekend, photos included.
• Online Shop AI Mistakes That Hurt Trust and Conversions — Other AI slips that cost shoppers' trust.
• Paint Store Marketing: Colour Advice Posts and Trade Offers With AI — Colour-advice posts built from your own chart, a coverage sum, honest wording on photo matching and a trade-account email with the margin maths shown.
• AI for Cosmetic Formulation: What It Can Draft and a Chemist Must Check — What AI can draft for a small cosmetics brand (briefs, ingredient lists, batch sheets) and what a qualified chemist or safety assessor must still check.
• Free AI Product Photoshoot: What Free Plans Really Allow — What free AI product-photo plans allow, where the limits bite, how to test them in 30 minutes and when a paid plan costs less than your time.
• Remove Backgrounds From Product Images With AI: Glass, Hair, Shadows — One-click background removers fail on glass, fur and shadows. Here is how to shoot, remove, repair and check edges, then batch the rest.
• AI for Skincare Products: Six Jobs a Small Brand Can Delegate — Six brand jobs for a small skincare label to hand to AI, each with a prompt, an example and the check that keeps claims safe.
• Can AI Write a Cosmetic Safety Report? What an Assessor Signs — What AI can prepare for a cosmetic product safety report, what only a qualified assessor can sign, and how to brief one so the job goes faster.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: CAP/ASA advice, Beauty and Cosmetics: The use of production techniques (asa.org.uk, dated 18 August 2025); Shopify help centre page on adding alt text to product media. Both checked 5 November 2026. The shade figures and returns sums are illustrative.