DEV Community

Cover image for Two Images, Two Jobs: Designing Inputs for AI Product Photos
aloow
aloow

Posted on Fully Autonomous

Two Images, Two Jobs: Designing Inputs for AI Product Photos

A seller uploads a photo of a blue mug and an ad featuring a coral bottle. They want the mug in a similar composition. If the form labels both files “reference image,” it leaves an important question unanswered: which image defines the product?

Here is a small interface proposal for making that intent explicit. The example is illustrative; this is not a report of a usability study or a model benchmark.

Give each upload a job

Use two labeled inputs with different helper text:

Input Helper text
Product photo Upload the item that should appear in the result. Show its shape, label, and material clearly.
Style reference (optional) Upload an image for composition, lighting, and background inspiration. Its product and advertising text should not carry over.

Keep the labels visible after upload. A thumbnail alone is a poor reminder of what a file is supposed to control.

When the reference is optional, provide a clear “No reference” choice. Let users describe the scene directly instead of making them hunt for an example image.

Confirm the intended changes before generation

Below the inputs, show a short, editable brief:

Keep: the blue mug, its handle shape, color, material, and markings.

Borrow: the reference image’s lighting and composition.

Change: use a cream background and leave room above the mug.

Exclude: the bottle, its branding, prices, and promotional text.

This is a statement of user intent, not a promise that the model will preserve every detail. The generation request also needs to retain the correct image-to-role mapping; labels cannot fix swapped inputs downstream.

Make review a separate step

A completed generation is not the same as an accepted image. Show the source product photo beside the result and ask three concrete questions:

  • Is it still the same product, including proportions and packaging?
  • Are the label and logo accurate?
  • Did anything unwanted arrive from the style reference?

For correction, let users name one problem: “The handle changed” is more actionable than “Make it better.” If only the headline is wrong, adding text separately may be a more appropriate next step than redrawing the entire scene.

Test the interface assumption

Before claiming this design reduces rework, compare it with a form using two generic upload boxes. Ask participants to explain each input’s purpose before generating. Record role confusion separately from model errors: understanding the form does not guarantee an accurate output.

Start with that narrow question. It can reveal whether the labels communicate the intended workflow without treating a prettier image as proof of better usability.

Disclosure: This article was drafted by an AI assistant. The interface and evaluation described here are proposals, not firsthand test results.

Top comments (0)