How to Edit a Photo with an AI Prompt (2026)

10 copy-ready edit prompts

Each prompt is a change line plus a keep line. Replace the specifics with your own photo's details. These are text templates; we have not attached before and after images to them.

Prompt anatomy: a change line and a keep line

The change line names the target, its location and the result: 'Change the jacket from black to dark green.' Location words (left, background, on the desk) matter because editors decide which pixels to touch from the text. Describe the result, not the process: 'a clear blue sky' works better than 'use a sky replacement filter'.

The keep line is the part most people skip. Image models regenerate more of the picture than you asked for, so naming what must stay (the face, the label text, the framing, the light direction) reduces drift. Keep both lines short; one change per prompt is easier to control than five.

Step by step

The same seven steps work in any prompt-based editor:

• Step 1: Start from the best photo you have. Use the sharpest, best-lit, full-resolution version. AI editors keep what is there; they do not reliably recover detail that is missing.
• Step 2: Write one change line. Say what to change, where it is in the photo, and how it should look afterwards, in one plain sentence.
• Step 3: Add one keep line. List what must not change: faces, expressions, text and logos, framing, lighting. Models change less when you name what to protect.
• Step 4: Pick a model and generate. Run the prompt once, then run it again or on a second model. The same prompt can give different results on different models.
• Step 5: Compare against the original at full size. Check the keep line first (faces, text, edges, framing), then whether the change looks right.
• Step 6: Fix one thing at a time. If something is off, edit the result with a short follow-up prompt about that one thing, or go back to the original and adjust the prompt.
• Step 7: Download at the size you need. Choose the output resolution before the final run so you do not have to upscale afterwards.

How to compare results

Open the original and the edit side by side at full size. Check the keep line first: are faces, hands and expressions the same person; is text still spelled correctly; did the crop or aspect ratio shift; do edges around hair and objects look natural? Only then judge the change itself.

Run the same prompt two or three times or on two models before deciding. If one model keeps identity and another gets the change right, it is often faster to switch models than to rewrite the prompt.

Which model to pick

Different models are good at different edits, so model choice matters as much as wording. In our own same-photo tests on October 9, 2026 (see Compare AI models), GPT Image 2.5 and Muse Image kept the original framing and pose most closely on a background swap and a relight, while both Nano Banana models reframed the portrait; for a simple text label all four did the job.

A practical rule: start with a fast, low-cost model for simple removals and recolors, and move to a stronger model when text, faces or fine product detail must survive. Check the model comparison page for side-by-side results before choosing for a large batch.

Common failure modes and fixes

What usually goes wrong, and what to try:

• The face changes: add 'keep the face, expression and identity unchanged' to the keep line, or try a model that preserved identity better on your photo.
• The whole image shifts or recrops: say 'keep the original framing and aspect ratio', and set the aspect ratio in the tool if it has one.
• Text or logos get garbled: name the exact text in the keep line ('keep the label text "Olive & Co" exactly'), and check it letter by letter.
• The edit spills beyond the area: add a location ('only the sky above the roofline') or use the tool's selection or brush if it has one; ChatGPT's help notes that highlights are not always precise.
• Lighting does not match: ask for the new element to match the existing light direction and color.
• Too many changes at once: split the prompt into separate edits and run them one after another.

Where Canva, Firefly, ChatGPT and Gemini fit

All of them accept the same change-plus-keep prompt. They differ in where the edit happens:

• Canva Magic Edit: click an element or brush an area, then type the change. Canva says it is powered by Google's Nano Banana model. Best when the photo is going into a Canva design.
• Adobe Firefly: Image, then Edit image; pick an Adobe or partner model (such as Nano Banana or GPT Image) and type the edit. Partner models use generative credits.
• ChatGPT: upload a photo and describe the change, or select part of the image with the selection tool first. OpenAI notes edits may extend beyond the selected area.
• Gemini: choose Create images, upload a photo and describe the edit. Google says Gemini images carry a visible watermark and an invisible SynthID mark.
• AI Media Studio (our product): upload, describe the change, and pick Muse Image, Nano Banana 2.1, Nano Banana Pro or GPT Image 2.5 per edit; Studio output at 1K, 2K or 4K, without a watermark.