Generating an image from scratch is one workflow. The other, often more common in marketing, is taking an image you already have and changing one specific thing about it. That's editing, and it's where AI tools save the most time: a product hero with the wrong background, a catalog shot with a discontinued accessory in it, a team photo that needs a former employee removed (make sure the reason is defensible and the edit doesn't misrepresent the team), a lifestyle photo with a distracting object. But editing also concentrates the risk. The closer an edited image looks to a real photograph, the more carefully you have to handle what it says, what it implies, and how you publish it — because in marketing, a public image is a claim.

Before you write any instruction, it helps to know what's on the menu, because the right vocabulary gets you a faster, cleaner result. Here's a table of the edit types AI can help with:

Most marketing edits are some combination of these. A "fix the product hero" request is usually object removal plus a background change plus a light color match. Naming the edit type in your head first makes the instructions you write next much sharper, because each type has its own failure modes (warped hands in inpainting, ghosted edges in object removal, seams in outpainting) that you'll want to call out as things to avoid — and in marketing, inpainting that invents a product detail you don't actually sell is also an accuracy problem.

Free-form editing requests like "remove the old packaging and update the background" are how good product shots get ruined. The model has no idea what you consider untouchable, so it'll happily redraw the product, shift a label, or invent a feature nobody approved. The fix is the Image Edit Brief, a four-part instruction:

• Preserve: what must remain exactly as-is.
• Change: the specific edit, described concretely.
• Avoid: artifacts, drift, or unintended modifications.
• Output: final format, dimensions, and where it's going.

Preserve is the part people skip. You have to explicitly name the product shape, the label, the logo, the colors, the lighting direction — anything you're not willing to lose, especially anything that affects accuracy. Change should be one or two surgical edits, not a wish list. Avoid catches the predictable failure modes: warped product, regenerated label text, invented features, mismatched lighting. Output ties the edit to its real use: a 1080x1080 feed post has different demands than a wide landing-page hero or a print spec.

Let's look at what a marketing edit conversation might sound like:

• Ryan: What's the edit?
• You: Take the old packaging out of the product hero and freshen up the background.
• Ryan: That'll come back with a product that looks slightly different than what we actually ship. What are you preserving?
• You: The product shape, the label text, the logo, the brand colors, the lighting direction. The product stays pixel-for-pixel.
• Ryan: Good. And Avoid?
• You: No regenerated label, no invented features, no seam where the packaging was, no off-brand background color.

Notice Ryan's pressure: name what stays before you name what changes, and in marketing, "what stays" includes anything that keeps the product honest. That order is the whole brief.

A clean edit doesn't automatically mean a publishable asset. Before any AI-generated or AI-edited image goes out under the brand's name, run it through five checks:

• Brand: does it match the visual system, or does it read as generic stock?
• Audience: will the people seeing it interpret it the way you intend, on this channel?
• Accessibility: does it need alt text, and is the contrast high enough where a headline or CTA overlays it?
• Accuracy: does the image imply something untrue — a real customer, a real result, a product feature or claim you can't support?
• Disclosure: if the image is AI-generated or substantially AI-edited, do the platform's rules, ad policies, or your audience's expectations call for labeling it (for example "illustration," "dramatization," or "not a real customer")?

The accuracy and disclosure checks are where marketing teams get burned. An AI image of "customers using our product," placed in a feed with testimonial-style copy, reads as a real endorsement even if you never used the word — an implied testimonial, and a rights problem if the faces resemble real people. A retouched product shot that quietly adds a feature becomes a false claim. The move when a risk surfaces is not always to scrap the asset: sometimes you revise the prompt, sometimes you change the caption or placement, sometimes you add a disclosure line, sometimes you swap in real photography or licensed UGC. The judgment call is yours, but it has to be a call, not a default.

The throughline of this unit: edits and AI images need a stated intent (the brief) and a stated standard (the five checks), or you'll publish something nobody actually approved — and in marketing, "published" means it's a claim. Before this becomes real, you'll get on a call with a brand lead who wants to ship an AI image tomorrow and isn't sure the risk is worth a delay. Your job in that conversation is to name the specific concerns, propose options, and land on a defensible call together: revise the prompt, change the usage plan, or kill the asset. Bring specifics. Vague risk talk loses to a publish deadline every time.