DEV Community

Cover image for Editing Photos by Describing the Change: A Practical Guide to Image-to-Image AI
Nick Chou
Nick Chou

Posted on

Editing Photos by Describing the Change: A Practical Guide to Image-to-Image AI

Most photo editors ask you to learn layers, masks and brushes. Image-to-image AI editors flip that around: you upload a photo, write what should change and what should stay, and the model returns an edited version. This post covers how that workflow works and the prompt habits that give the most predictable results.

What "image-to-image" editing means

A text-to-image model starts from nothing. An image-to-image editor starts from your photo and applies an instruction to it, so the subject, composition and lighting carry over unless you ask for them to change. That makes it a good fit for everyday retouching jobs:

  • Restoring and colorizing faded or torn family prints
  • Removing tourists, cables or other distractions from travel shots
  • Swapping the background of a product photo for a clean listing image
  • Changing season, weather or time of day
  • Restyling a portrait as anime, Ghibli-like or 3D-cartoon art
  • Previewing a new hair color or outfit

Prompt habits that help

  1. Say what to keep, not only what to change. "Change the background to a white studio backdrop. Keep the product, its shadows and its colors exactly the same."
  2. Change one thing per request. Chain several small edits instead of one long instruction; if a step goes wrong, you only redo that step.
  3. Be concrete. "Remove the two people on the left and fill in the stone wall" works better than "clean this up".
  4. Name the output you want. "Square crop, soft daylight, no text" tells the model what finished looks like.
  5. Start from a decent source. Sharp, well-lit JPG or PNG files (up to 10 MB in most tools) give the model more to work with than heavily compressed screenshots.

Trying it with NanoPic

NanoPic is an image-to-image photo editor built around a chat interface. You upload a photo, describe the edit, and the result appears in the conversation, where you can download it or refine it with a follow-up instruction. You can choose between GPT Image 2/2.5 and Nano Banana 2/2.1/Pro for each request, and output goes up to 4K. New accounts get 10 free credits without a card, and there is a public showcase gallery of ready-made prompts you can copy and adapt.

A quick first experiment: take an old, faded photo, and ask for "Restore this photo, remove scratches and creases, add natural color, keep the faces unchanged." Then follow up with a second instruction to adjust brightness or crop.

Limits to keep in mind

  • Faces and fine text can still drift on aggressive edits, so keep instructions narrow when identity matters.
  • Always check the result at full size before using it commercially.
  • Check the license terms of whichever tool you use before publishing edited images.

If you have favorite prompts for restoration or product photos, share them in the comments.

Top comments (0)