The short version: the best ChatGPT prompts for image editing name the exact change you want, the part of the image it applies to, and what must stay untouched, in that order. Vague prompts like “make this better” burn your time and your monthly image credits. The prompt matters less than most guides claim; the quality of your starting photo and your willingness to correct in rounds matters more.
Why most people pick the wrong prompt first
I watch this happen constantly with clients. Someone opens ChatGPT, drops in a product photo, and types something like “improve this image” or “make it look professional.” Then they wait, get something odd back (a hand with six fingers, a logo that’s melted, a background that’s still cluttered), and conclude the tool is rubbish.
It isn’t rubbish. The prompt was the problem. ChatGPT’s image tools (built on OpenAI’s image generation model, the one that shipped inside ChatGPT in March 2025 and has been quietly updated several times since) respond well to instructions that read like you’re briefing a junior designer over Slack, not like you’re casting a spell. A junior designer needs to know what to change, what to leave alone, and what “done” looks like. So does the model.
The five things a good image-editing prompt needs
Before you type anything, get these five pieces sorted in your head. I keep this as a mental checklist now and it’s cut my revision rounds roughly in half.
- The specific edit. Not “fix the lighting” but “brighten the left third of the image where the shadow falls across the product.”
- What must stay identical. Faces, logos, product shape, exact colour codes. If you don’t say “keep the packaging text unchanged,” expect it to change.
- The style reference. “Match the warm, slightly desaturated tone of a film photo” gives the model something to aim at. “Make it nicer” does not.
- The output format. Square for Instagram, 16:9 for a website banner, transparent background for a logo. State it, don’t assume it.
- The failure you’re trying to avoid. Sounds odd, but telling it “don’t distort the hands or the text on the sign” cuts down on the garbled results.
If you want a wider vocabulary for describing tone and style precisely rather than reaching for generic words like “nice” or “professional,” this list of descriptive words for business prompts is worth keeping open in a tab. Specific adjectives do a lot of the heavy lifting in image prompts too, not just text ones.
My Tuesday-afternoon headshot disaster
Last spring I needed a new speaker headshot for a conference deck, quickly, and I didn’t want to book a photographer for one image. I took a decent phone photo against my kitchen wall and uploaded it to ChatGPT with the prompt “make this look like a professional studio headshot.”
Forty minutes and eleven versions later, I had a photo of a woman who looked like me from a distance but with a jawline that shifted slightly every single time, a blazer that kept changing colour between navy and black, and a background that once included a plant that was never in the original photo. I’d given the model nothing to hold onto.
Version twelve worked because I rewrote the prompt like this: “Replace only the background with a plain, softly blurred charcoal grey studio backdrop. Do not alter my face, hair, skin tone, or expression. Keep the blazer the exact navy colour shown in the original. Add soft, even studio lighting from the front left, similar to a corporate LinkedIn headshot. Crop to a square 1:1 format.” One pass. Done.
The lesson wasn’t really about wording cleverness. It was that I’d finally told it what to leave alone. That single line, “do not alter my face,” was doing more work than the rest of the prompt combined.
A simple process for choosing the right prompt
Here’s the process I now use with every image edit, whether it’s for a client’s product shot or my own social content:
- Decide the single job. One prompt, one edit. Don’t ask it to remove the background, change the lighting, and swap the outfit all at once. Split it into three prompts, three passes.
- Write the “keep unchanged” line first. Before you even describe the edit, decide what’s off limits and put it in the prompt explicitly.
- Reference something real. “Like a magazine editorial shot” is weaker than “like the lighting in an Anthropologie catalogue photo” or “flat, even lighting like a Shopify product listing photo.” Concrete references beat abstract adjectives every time.
- Ask for one output, not options. If you ask for “a few variations,” you’ll get inconsistent quality across all of them. Ask for one, judge it, then refine.
- Correct, don’t restart. If it’s close, don’t rewrite the whole prompt. Add a follow-up: “keep this exact version but reduce the shadow under the chin.” That short conversational correction usually gets you there in one or two more turns rather than starting the whole prompt from scratch.
- Set a limit before you begin. I now give myself five rounds maximum for any single image. If it’s not right by round five, the starting photo is the problem, not the prompt.
Prompts by edit type (with real examples)
Background removal or replacement
Weak: “remove the background.”
Stronger: “Remove the background entirely and replace with a solid white (#FFFFFF) background, keeping the product’s shadow beneath it soft and natural, as you’d see on an Amazon listing photo.”
Colour correction
Weak: “make the colours pop.”
Stronger: “Increase saturation slightly on the red and orange tones only, leave skin tones and whites unchanged, similar to a warm autumn Instagram filter.”
Object removal
Weak: “clean this up.”
Stronger: “Remove the coffee cup in the bottom right corner and fill the space naturally with the same wooden desk surface visible elsewhere in the photo.”
Style transfer
Weak: “make it artistic.”
Stronger: “Apply a flat, minimal illustration style with muted pastel colours, similar to a modern editorial magazine cover, while keeping the layout and composition identical.”
Notice the pattern. Every stronger prompt names a real, specific reference point a human designer would understand instantly. The model works the same way a freelance designer does when you give them a brief; vague briefs get vague work back regardless of the tool. I cover this same pattern-matching problem across other AI tools in AI tools for freelancers, because it isn’t unique to images.
The bit most prompt lists get wrong
Here’s something the “50 best ChatGPT image prompts” listicles rarely admit: copy-pasting someone else’s prompt from a blog post or a Twitter thread mostly doesn’t work the way it did when they posted it. OpenAI updates the underlying image model quietly and fairly often, and a prompt that produced a crisp result in June can produce something warped or over-stylised by August because the weighting behind certain phrases shifted. I check weekly AI news roundups partly for this reason, because image tool behaviour drifts between updates, and last week’s viral “magic prompt” is this week’s disappointing one.
The uncomfortable part is that this means there’s no permanent perfect prompt to memorise. What stays consistent isn’t the wording, it’s the structure: name the edit, name what stays fixed, give a real reference, limit the scope. That structure has held up for me across every model update in the last year even when specific phrasing has needed tweaking.
There’s also a cost angle worth being honest about. If you’re doing this at any volume, say editing product photos for an ecommerce catalogue or headshots for a whole sales team, ChatGPT Plus at roughly £16 to £20 a month has generous but not unlimited image generation, and heavy use pushes people toward the API or a dedicated image tool instead, which changes the maths. I’ve broken this down in AI tool costs explained for business owners if you’re trying to work out whether ChatGPT is even the right tool for the volume you need.
When to stop prompting and start over
A blunt rule I’ve adopted: if you’re on your sixth or seventh attempt and the core problem (a distorted face, a logo that keeps mutating, text that won’t render correctly) hasn’t shifted, stop. The photo itself is likely too low resolution, too dark, or too cluttered for the model to work from cleanly. No prompt fixes a bad source image. Go back, take or source a better original, and start your five-round limit again from there.
Once you’ve got a batch of edited images you’re happy with, where they land matters as much as how they were made. A gorgeous background swap does nothing for you sat unused on your desktop. If the end goal is social content, it’s worth thinking about which platform and scheduling setup gets them in front of people; I compared the practical options in the best social media post tool for small teams.
Related reading: chat gpt portrait.
Related reading: ai photo editing strengths.
If your question is a different ChatGPT prompts one, the ChatGPT prompts guide lists every answer I have written.
Frequently asked questions
What makes a ChatGPT image editing prompt good rather than average?
A good prompt states the single specific edit, names exactly what must stay unchanged (face, logo, colours, text), and gives a real style reference rather than a vague adjective like “professional” or “nice.”
Why does ChatGPT keep changing small details I didn’t ask it to change?
Because if you don’t explicitly tell it what to preserve, it treats the whole image as fair game for reinterpretation. Add a firm “do not alter” clause covering faces, logos, or exact colours and this drops sharply.
How many attempts should I expect before an image edit looks right?
Budget for up to five rounds of refinement on a single image. If it’s not close by round five, the source photo (resolution, lighting, clutter) is usually the real problem, not the prompt wording.
Are the ChatGPT prompt lists shared online still reliable?
Treat them as a starting structure rather than a script. OpenAI updates the image model periodically, so exact wording that worked months ago can behave differently now, even though the underlying principle of naming the edit and what to preserve still holds.
