How to Brief an AI Image Model for On-Brand Product Photos
Five prompts for the same product can produce five different brands. Here is how to write a reusable brief that keeps AI product photos consistent, accurate and on-brand.
Some links in this post are partner or affiliate links. Unified WebTech is a DashNex Authorized Partner and may earn a commission if you sign up, at no extra cost to you. See our disclaimer.

On this page
Ask an image model for "a photo of our candle on a wooden table" five times and you get five different brands. One is moody and dark, one is bright Scandinavian minimalism, and one has a lens flare nobody asked for. Each image is fine alone. Side by side on a product page, they look like five different shops.
An AI image model brief for product photos fixes this. It is a written document that covers the same inputs every time, so the model has less room to improvise. Below is how we build one, how to write the per-product prompt, and how to check what comes back before customers see it.
Why most AI product photos look off-brand#
A vague prompt allows a huge range of acceptable answers, and the model picks from that range differently on every run. Rewording alone rarely narrows it. You just land in a different random corner.
Photographers do not work this way. A studio shoot has a fixed light setup, a fixed backdrop, a fixed lens and a shot list. Variation comes only from the product. Your AI workflow needs the same discipline, written down instead of set up on a tripod.
So we stop hunting for better wording and write a brief. Some inputs stay fixed across every image. A few change per product. With the fixed part pasted in every time, we see far fewer off-brand results, and the ones that remain are easier to diagnose because you know what was held constant.
Build a brand style sheet the model can reuse#
The style sheet is the fixed part. Write it once, keep it to one page, and paste it into every prompt. It covers six things: palette, lighting direction, lens feel, background rules, mood words, and reference photos you already like.
Colors, light and background#
Use hex codes for color, not names. "Warm cream" means something different to every person and every model. #F3EBDD is far less ambiguous. Models still match hex codes only approximately, so you will verify the result later with an eyedropper.
A sample block:
PALETTE: background #F3EBDD, accent #2F4A3A, shadow tone #5A4A3A.
No other saturated colors in the scene.
LIGHT: soft window light from camera left, gentle fall-off to the right,
shadows soft-edged and warm, no harsh flash, no rim light.
LENS: 85mm look, shallow depth of field, product sharp front to back,
background gently blurred, no wide-angle distortion.
BACKGROUND: plain seamless paper or matte plaster surface, no text,
no logos, no clutter, horizon line low and level.
Then add two to five photos you already like: past shoots, competitor shots you admire, magazine references. Attach them if the model accepts image input. If it does not, describe what you like about each in one sentence. "Light skims the left edge of the jar and leaves the label in soft shade" is far more useful than "nice lighting."
Words to use and words to avoid#
Mood words steer the whole image. Pick three to five you will always use, and a short list you will never use.
- Use: calm, tactile, natural, understated, warm.
- Avoid: luxurious, dramatic, glossy, cinematic, hyper-realistic, epic.
Avoid-words matter because loud adjectives can push a model toward heavy color grading, flares and gloss. If your brand is quiet, ban them in writing. Also list visual avoids: no lens flare, no floating objects, no stock-photo smiles, no text overlays.
Keep the list honest. If you are copying someone else's adjectives, check them against your five best-selling products. The words should describe what already works.
Write the per-product prompt#
The style sheet stays fixed. Everything else changes per product.
The five-part template#
We use five parts, in this order:
- Subject. What the product is, in plain words.
- Product details. Size, material, color, finish, and any visible features. Be literal.
- Camera angle. Eye level, 45 degrees from above, top-down, or close-up on a detail.
- Surface and props. What it sits on and what else is in frame. Keep props to two or fewer, and name them.
- Style sheet. Paste the whole block unchanged.
Product details come before style. Many models seem to give more weight to earlier words, and product accuracy matters more than mood. Treat that as a habit worth testing, not a rule.
A worked example#
Take a made-up product, an 8 oz soy candle called Cedar & Fig:
SUBJECT: A single soy candle in a straight-sided amber glass jar
with a matte black lid, lid removed and set beside the jar.
PRODUCT DETAILS: Jar about 8 cm wide and 9 cm tall. Cream paper label
on the front, blank area only, no readable text. Wick lit, small
steady flame. Wax surface smooth and pale ivory.
CAMERA ANGLE: Eye level, slightly off-center, product fills about
half the frame height.
SURFACE AND PROPS: Matte plaster surface in warm off-white.
One sprig of dried cedar to the right. No other props.
[Paste style sheet: palette, light, lens, background, mood words,
avoid list]
Look at the label. We asked for a blank one, because image models often garble printed text. We add the real label afterward, in an editor or by compositing the actual label artwork.
Here is our one firm opinion. If your product has packaging with text, a logo or a distinctive shape, do not ask the model to invent it. Start from a real photo of the real product and use AI for the scene around it, or generate the scene and place the product in. Generating the product from scratch is how you end up with a jar that does not exist.
Keep a catalog consistent with reference images#
One good image is easy. Forty that look like one shoot is the real job.
Use a reference image per product line. Pick the best approved image from a line, such as all candles or all tote bags, and attach it to every later prompt in that line with an instruction like "match the lighting, surface and framing of the reference." If your model supports seeds or style references, record the setting next to the prompt.
Use the same settings for every item. Same aspect ratio, same angle, same crop, same background. Change only the product. Write the settings into the brief so a teammate does not have to guess. Mixed aspect ratios make a product grid look sloppy faster than anything else.
Know when to regenerate and when to edit. Our rule:
- Regenerate when the composition, lighting or product shape is wrong. Those problems are structural, and patching them rarely works.
- Edit when the image is right except for a small flaw: a stray highlight, a blotchy edge, a wrong shade, a garbled label. Use inpainting or a normal photo editor.
Regenerating a nearly perfect image to fix one detail usually breaks something else. Edit it.
Check every output before it goes live#
AI images fail in ways a normal photo does not. Run each one through this checklist. For anything customers are likely to return if it looks different from the picture, have a second person run it too.
- Product accuracy. Compare against the real item. Check proportions, color, finish, number of parts, closures and handles.
- Labels and text. Zoom in. Any readable text should be correct, or absent. Warped letters and invented brand names are common.
- Hands and edges. If a person appears, count the fingers and check the grip. Look at where the product meets the surface for melted edges, odd shadows or fused objects.
- Brand colors. Sample the background with an eyedropper and compare it to your hex codes. A few shades of drift is normal. A visible shift is not.
- Implied features. Does the image suggest a waterproof finish, a larger size, extra accessories or a function the product lacks? Remove it or regenerate.
Log failures as you go. If three images in a row show a warped lid, add a line to the brief: "lid is flat and perfectly circular."
Disclosure, rights and where AI photos mislead#
Settle three things before publishing.
Platform terms. Read the commercial-use terms of your image tool, and read them again when you change plans or tools. Terms differ on who owns outputs, whether commercial use is allowed on every tier, and what happens to reference images you upload. Keep a copy of the terms as they stood when you generated the images.
Disclosure. Some markets and marketplaces expect you to say when imagery is AI-generated, and the rules are changing. Check them for each place you sell. Even where nothing is required, a short note in your image policy is cheap and protects trust if a customer asks. Never present a generated image as a photo of a specific customer or a real event.
Misleading images. The risk we see most is quiet misrepresentation, not outright fakery. A candle shown larger than it is, a fabric that looks thicker, a before-and-after that implies results the product cannot promise. If an image makes a customer expect something the product does not deliver, it is a refund request waiting to happen.
Product images are one asset in a bigger system. Our post on an AI content pipeline for small business shows how images, copy and short video can come from one script. Once your images are approved, run them through the website launch checklist for file sizes, alt text and page checks.
FAQ#
How do I choose an image model?#
Test two or three on the same prompt, with your style sheet pasted in, using one product you know well. Judge by product accuracy and how closely the output follows your palette, not by which looks most impressive. A multi-model tool such as GenX Pro can make side-by-side comparison easier, but separate accounts work too. Pick the model that needs the fewest corrections.
How many revisions should I expect per product?#
Expect several rounds on the first product in a line while you tune the brief, and fewer after that. Once the style sheet and reference image are settled, many products need only a few tries plus small edits. We would not promise a number. Track your own average over the first ten products and use it to budget time.
Do AI photos suit marketplaces with strict rules?#
Sometimes. Many marketplaces set specific rules for main images, such as background, framing and accurate depiction of the item. Read those rules first. A common safe approach is real photos for the main listing image and AI-generated lifestyle scenes for secondary images or your own site, as long as they stay accurate.
How should we store briefs so the team reuses them?#
Keep one shared document per brand, with the style sheet at the top and a section per product line holding its reference image, approved prompt and settings. Add a short "known failures" list. Name files with the product and date, and keep the approved prompt next to the approved image. If nobody can find the brief in thirty seconds, they will write a new one.
Your next step this week: write the one-page style sheet, run it on your best-selling product, and check the result against the five-point checklist. Fix the brief, not the output, and you will have a repeatable process before the second product.
Get new posts by email
One email when something new is published. Unsubscribe any time.

