AI Image Editing: Which Tool for What Job

Most AI image work is editing, not generating. A product shot needs a cleaner background, a portrait needs a wider crop for a banner, a campaign image has one wrong detail, a generated poster has a misspelt word. The tools for these jobs now fall into three families: selection-based editors where you paint or select an area, one-click utilities for routine jobs such as removing a background, and instruction editors where you describe the change in plain language. This guide sorts the main tools by the edit you need to make.

The tools at a glance

Features move between free and paid plans, and partner models usually cost more credits. Check the current plan before relying on any one tool for client work.

Remove, replace and add (inpainting)

For removing a stray object, Photoshop’s Remove tool and Canva’s Magic Eraser are the fastest: brush over it and the area is filled. For replacing or adding something in a specific area, Generative Fill in Photoshop lets you select the area, type what should be there, and choose the model, including Adobe’s Firefly models or partner models from Black Forest Labs and Google. Canva’s Magic Edit does the same inside a design and is powered by Google’s Nano Banana. Midjourney’s V8.2 Edit Model brings inpainting into Midjourney itself.

The rule for all of them: select a little more than the object, including its shadow, and describe what should be there, not what should go.

Expand the canvas (outpainting)

Turning a portrait image into a landscape banner, or a square into a 9:16 story, is now routine. Generative Expand in Photoshop and Magic Expand in Canva extend the background in any direction, and the Midjourney Edit Model can expand the canvas too. Photoroom offers AI Expand for switching product images between formats. Expect the new area to need one or two attempts when it contains people or text. For the Canva tools, see Canva Magic Studio Edits.

Backgrounds

Removing a background is a one-click job in Photoroom and Canva. Photoroom’s model is trained on product photography and can process batches, which matters when you have a catalogue rather than one image. Generating a new background is where the tools differ: Photoroom creates scenes from a prompt or a reference image and adds AI shadows, while Photoshop’s Generate Background keeps you in a layered file. Product-specific advice is in ChatGPT Images for Product Mockups.

Describe the change (instruction editing)

Instruction editors are the biggest shift of the last year. You upload an image and say 'make it evening', 'swap the jacket for a navy coat' or 'remove the people in the background'. Nano Banana 2.1 is built for multi-turn editing and, according to Google, improves character consistency across turns; see Gemini Image Editing Guide. ChatGPT Images 2.5 adds comments placed on part of an image so only that area changes, plus a sketch tool for rough guides; see ChatGPT Image Edits and References. FLUX Kontext is known for literal, predictable edits that leave the rest of the image alone; see Flux Kontext Editing Guide.

Ask for one change per instruction. Stacking five changes in one message is the most common reason an edit drifts. When an edit chain works, keep it: Dreamdrive saves each version with the prompt that produced it, so you can find the exact step where the image went right.

Text, light and resolution

For a misspelt word, regenerate the text in a tool known for typography. Ideogram is the usual choice, covered in Ideogram Text and Typography, and ChatGPT and Nano Banana both list better text rendering in their 2026 updates. For relighting, Magnific Relight is known for changing the light on an existing image. For final resolution, an upscaler is a separate step; compare them in Best AI Image Upscalers Compared.

Try it

Take one photo of a person outdoors and run the same three edits in two tools: an instruction editor and a selection-based one.

1. Remove the lamp post on the left.
2. Change the time of day to golden hour, keep everything else the same.
3. Extend the image to a 16:9 banner with more sky on the right.

Compare three things: did the person’s face and clothes stay identical, did the removed area look natural, and did the extended sky match the original light? Then note which tool you would trust with a client file. Write down the model each tool used, because several of them now let you switch models, and results differ between them.