Editing Images by Instruction and Reference

Most real image work no longer stops at the first generation. You generate, then edit: make the jacket green, remove the lamp post, put this product in that scene, keep this character the same across ten frames. Current models do this by instruction, and the instruction is a prompt with its own rules. Start by predicting what a vague one does.

Editing is now the main workflow

As of November 2026, all the major image tools edit. OpenAI's GPT Image models modify existing images from a new prompt, partly or entirely; masking is prompt-based, they accept several input images, and the Sunburst variant is tuned for editing precision. Google's Nano Banana 2.1 takes up to 10 object images and 4 character images as references. FLUX.2 [pro] takes up to 8 reference images. Midjourney's docs describe a V8 Edit Model that replaces its earlier Omni Reference, Character Reference and Retexture features and takes up to 4 reference images.

The edit instruction has two halves. The change: one specific thing, located precisely (“the coffee cup in her right hand”). The preserve list: everything that must not move (“her face, expression, pose, clothing, the window light and the background”). The preserve list feels redundant, since the model can see the image, but it's what stops the model from improving things you never asked it to touch.

Locate edits in words. Most edit-by-instruction models don't need a painted mask; GPT Image's masking, for example, is prompt-based. Describe the region the way you'd describe it to a retoucher on the phone: “the lamp post at the far left edge”, “the second shelf from the top”, “her left sleeve”. For a text change, put the new wording in quotation marks and say which existing text it replaces.

Edit the attached photo.

Give every reference a role

With several images attached, the model has to guess which one supplies what. Don't make it guess. Number the images and name each one's role, the way BFL's FLUX.2 guide suggests: “subject from image 1, style from image 2”. Say which details of the subject must survive (the logo, the stitching, the exact colors), and describe the result as one scene rather than a list of operations.

Iterate in small steps

Make one change per turn and check it before the next. Drift accumulates over a long chain of edits: a logo softens, a face shifts. When it does, go back to the last good version, re-attach the original reference and restate the preserve list, rather than piling corrections on a degraded image. For a set of images that must match, such as one product in eight scenes, reuse the same reference image and the same preserve list every time, and change only the scene. Settings matter too: OpenAI notes that higher complexity can help with dense text and complex layouts but can also add latency and unwanted changes, so compare results for your task.