gpt-image-edit

Run 'runcomfy run openai/gpt-image-2/edit' with a JSON input containing a prompt and an images array of HTTPS URLs. Lead the prompt with preservation goals like keeping the face and brand unchanged, then state the edit.

GPT Image Edit is strongest for multilingual in-image text, identity preservation, and layout-precise edits with up to 10 references. Nano Banana Edit handles batches up to 20 images, and Flux Kontext suits single-shot source-fidelity-first local edits.

Quote the exact characters and name the script in the prompt, such as 'the headline reads "コーヒー" in bold Japanese kana'. Keep the rest of the image preserved by stating that layout, photograph, and brand mark stay unchanged.

The images array accepts up to 10 publicly fetchable HTTPS URLs. The first image is primary and the rest are auxiliary; refer to them by number in the prompt, such as 'subject from image 1, lighting from image 2'.

A 422 occurs when the size field is set to something other than auto, 1024_1024, 1024_1536, or 1536_1024. Use size auto to preserve the input ratio unless the edit explicitly changes framing.

Avoid it for batch consistency across many SKU images, where Nano Banana Edit handles up to 20 images better, and for portrait photorealism, where Nano Banana Pro wins. Long compound edit prompts also drift, so split them into multiple passes.