Skip to article
KnaccoGuides
Creative work

How to get consistent AI product images across a whole set

How to make product photos with AI that stay consistent: build a reference set, write an identity sheet, change one thing per request and check each image.

Knacco two characters comparing matching product bottles.
0:00
7:05

To make product photos with AI that stay consistent from one image to the next, anchor every generation to approved photos of the real product, describe its fixed details in the same words every time, and change only one thing per request. Then check each result against the original photo before you judge the scene. This keeps the cap, label and proportions the same across a whole set.

Most drift starts with a shortcut. You make one good image, then use it as the reference for the next. Each generation copies the last one's small errors, and by the fifth image the handle is thinner and the logo has moved. Always go back to the real product photos.

Build a small reference set

Four or five good photos of the actual product do more than twenty loose ones. Aim for:

  • Front view. The main reference for shape, color and logo position.
  • Side view. Shows depth, handles, lids and anything that sticks out.
  • Back or label view. Shows required text, ingredients or markings.
  • Close-up of a distinctive detail. A hinge, a stitch pattern, a texture.
  • Scale view (optional). The product beside a common object, for size.

Use the original files from the camera or the manufacturer. Screenshots and compressed web images lose the edges and fine detail you need. Name each file by what it shows, such as bottle-front.jpg and bottle-lid-closeup.jpg, so you can point to it in the rejection log.

If a side of the product is not in your reference set, do not ask for a view that shows it. The model will invent that side. Photograph it first or choose a different angle.

Write a product identity sheet

The identity sheet is a short, fixed description you paste into every request. Writing it once stops you from describing the product differently each time.

Product: [NAME, VARIANT, SIZE]
Shape: [E.G. TALL CYLINDER, ROUNDED SHOULDERS, FLAT BASE]
Material and finish: [E.G. MATTE STAINLESS STEEL]
Color: [NAME], matching the front reference photo
Logo: [POSITION AND SIZE, E.G. CENTERED, LOWER THIRD]
Must never change: [LID SHAPE, HANDLE, LABEL LAYOUT]
Must never be added: [EXTRA BUTTONS, TEXT, ACCESSORIES]

Keep the sheet with the reference files. When a teammate makes the next batch, they use the same words and the same photos.

Change one thing per request

Each request should keep the product block identical and change one element: the setting, the light or the camera angle. That makes the change easy to judge and easy to undo.

A working request looks like this:

"Use the attached front and side photos as the exact product. [PASTE IDENTITY SHEET]. Keep the product unchanged. Change only the setting: a plain oak desk beside a closed notebook, soft window light from the left. Leave clear space above the bottle for text added later. Do not add any text to the image."

When you want a new angle, keep the setting and light from the last approved image and change only the angle. Changing three things at once makes it impossible to tell which one broke the product.

Fix the look of the set the same way. Once a setting and light are approved, write them as one style line and paste it after the identity sheet in every request:

"Style for this set: warm gray background, soft light from the left, camera at product height, gentle shadow under the base."

The product block keeps the item the same. The style line keeps the images looking like one set.

Review the product before the scene

A polished scene can hide a small product error, so look at the product first. View each result at full size, beside the matching reference photo, in this order:

  1. Shape. Outline, proportions, lid, handle, base.
  2. Color. Compare in the same viewer, at the same brightness.
  3. Surface details. Seams, texture, hinges, reflections on the product.
  4. Text and logo. Spelling, position, size.

Only after the product passes do you judge the setting, mood and composition.

Keep a rejection log

Write down why each image failed. "Looks off" gives you nothing to fix. "Wrong lid shape" does.

Image Reference What is wrong Correction to request
desk-02 bottle-side.jpg Lid hinge missing Restore the hinge from the side photo. Keep the setting.
park-01 bottle-front.jpg Logo too high Move logo to the lower third as in the front photo.

Ask for one correction at a time and repeat the identity sheet each time. Keep the last approved version, so a correction never erases good work. If the same error appears in three images, improve the reference set or simplify the scene before the next batch.

Add exact text and logos in the layout

Small generated text is unreliable. It can look right at a glance and be misspelled at full size. If a label, legal line or price must be exact, add it afterwards from approved artwork in your design tool.

The same goes for crops. Export each image at the sizes you need, then check the exports, not only the master. A square crop can cut off the detail that made the product recognizable.

Mistakes that break product consistency

  • Using a generated image as the reference. Errors copy forward. Always return to the real photos.
  • Describing the product differently each time. Different words produce different products. Paste the identity sheet.
  • Changing several things at once. You cannot tell which change caused the error.
  • Judging the mood first. A beautiful scene makes a wrong lid easy to miss.
  • Trusting generated text. Check every word at full size, or add the text yourself.

Make the product block reusable

The identity sheet and the fixed request above are what keep a set consistent, and they are the part worth reusing. Prompts for Consistent Image Sets helps you build reusable prompt instructions that keep characters, products and visual style consistent across a set of images.

It installs as a skill in Claude Code, Codex or Antigravity CLI. You ask the agent for a four-shot set of your product, with your product photo as the reference, or you paste a prompt that keeps drifting and ask for a review. You then render once after each single change in your own image generator, which is billed separately. Check its requirements on the product page before you choose it.

For a start that costs nothing, Knacco's free skills and prompt packs include Product Photo & Video Prompts, which writes a matching image prompt and motion prompt from one product brief and checks the pair before you generate.

Prompts for Consistent Image Sets belongs to the Product Photography bundle, with the skills for coordinated product photos, prompt refinement and image review. The bundle costs less than its products bought one by one.

If you have not yet written down what the set is for, start with an AI creative brief template. Store the finished identity sheet with its references in your AI prompt library.

Before you publish the set

  • Every image was checked against a real product photo, not a generated one.
  • Shape, color and details match at full size.
  • Required text was checked word by word or added from approved artwork.
  • Each exported crop still shows the product clearly.
  • Accepted and rejected images sit in separate folders.

Common questions

Why does AI keep changing my product in images?

AI image tools change products because they redraw the whole image each time and fill missing details with plausible guesses. Drift gets worse when you use a generated image as the next reference or describe the product differently in each request. Use real product photos every time and paste the same fixed description into every prompt.

Can AI keep a logo or label accurate?

AI image tools often get logos and small text slightly wrong, even when the image looks convincing at a glance. Check every word at full size against your artwork. When a label, legal line or price must be exact, add it afterwards from the approved artwork file in your design tool.

How many reference photos do I need for AI product images?

Four or five good reference photos usually cover a product: a front view, a side view, a back or label view and a close-up of a distinctive detail, plus a scale photo if size matters. Use original camera or manufacturer files. Avoid asking for any view the reference set does not show.

Should I use an AI image as the reference for the next one?

No. Using a generated image as the reference for the next one copies its small errors forward, and they build up across a set. Always return to the approved photos of the real product. Generated images can guide the style of the set, but the product itself should come from real references.

Sources and further reading (2)
Back to all guides
Keep exploring

A good next read.