Preserving product details with GPT-image-1 input fidelity

The guide, rather than a new model

OpenAI published its high input fidelity cookbook on July 17, 2025. It documents input_fidelity="high" for preserving details during GPT-image-1 edits. The date here is the guide publication date, not a claimed model launch.

Define what must stay unchanged

Write an edit brief in two parts: the requested change and the protected details. For a product shot, you might change the background while protecting the bottle silhouette, label wording, cap color, and relative proportions.

Choose a clean source image. A compressed screenshot with small lettering is a poor reference for judging whether an edit preserves typography. Retain the original file and compare it with the result at a matching scale.

Review fidelity as a separate decision

Make a small paired evaluation with the same sources and instructions. Record whether each protected detail survives, then compare request cost and the amount of human correction required. Do not use visual polish alone as the acceptance criterion.

The historical guide also calls out reference ordering. Keep your source list stable during an experiment; changing both the references and a setting makes the comparison less useful.

Carry the method forward

Fidelity controls differ across model generations. Check the reference for your selected model before reusing an older parameter. The review method remains useful even when the request schema changes.