← Back to blog
TutorialsJul 14, 2026 · 6 min read

Image-to-image: keep the product, change everything else

Reference-based editing is the most underused feature in GPT Image 2. Here's how to reshoot a product across seasons, settings and styles without losing its identity.

I

ImageTwo Editorial

Workflow guide

Image-to-image: keep the product, change everything else

Text-to-image is a great party trick, but the moment you need the same object to appear consistently across a campaign, you need image-to-image. Upload a single clean product shot and you can re-render it in a winter scene, a summer scene, and a studio scene — with the product itself staying pixel-faithful.

Start with a clean reference

The model preserves what's unambiguous. Use a reference with a clear silhouette, even lighting, and no background clutter. A busy reference forces the model to guess which parts matter, and it will guess wrong at the worst moment.

Describe the change, not the object

The most common mistake is re-describing the product in the prompt. Don't — the reference already carries that. Instead, describe only what should change: "the same bottle on wet sand at dawn, soft fog, cool blue light". Every token you spend re-describing the bottle is a token that can drift its details.

Iterate in small passes

Resist the urge to change five things at once. Change the setting, review, then change the light, review. Small passes keep the identity locked and make it obvious which edit introduced a problem.

Put this into practice

Open the Studio and try it on your next image.

✦ Open Studio