ChatGPT's image tools just got usable for real work

Back in August I ran one of our own real estate ads through ChatGPT and wrote up the honest result. It got about ninety percent of the way there. The last ten percent was the maddening part: the description came back too bold, the address split onto two lines, the photo grew.

You could see exactly what was wrong. You just could not tell it that. Every attempt to fix one thing meant it redrew the whole ad and handed you three new problems.

On Tuesday, OpenAI shipped the fix. It is called ChatGPT Images 2.5, and it comes with four new things you can actually put your hands on.

This is not a small corner of the product, by the way. OpenAI says people now make more than three billion images a week across ChatGPT and its image tools for developers.

What actually changed

The plain version: sharper, faster, and much better at leaving alone the parts you did not ask it to touch.

Four claims from OpenAI worth knowing:

• Faster. Generation is up to fifty percent quicker than the previous version. In practice that is the difference between trying four ideas and trying eight.
• It changes only what you asked for. This is the big one. Edit the headline, and the photo, the layout, and the colors underneath it stay put.
• It holds together over a long back and forth. Ten rounds of small edits used to turn an image to mush. Each change now builds on the last instead of quietly degrading it.
• It is better with your own photos. Hand it a picture of a real person or a real building and what comes back actually looks like them.

It also handles transparent backgrounds and more complicated layouts than it used to, which matters more than it sounds like it does. More on that below.

The four new tools

Sketch

You can now draw directly inside ChatGPT and have it use your drawing as the plan.

Type @Sketch in ChatGPT and a drawing pad opens. Scribble the layout: box in the corner for the logo, big rectangle for the photo, three lines of text across the bottom. Then describe the style and the details in words, and it builds the finished image on top of your rough sketch.

Nobody is grading your artwork. The point is that a thirty second doodle tells it where things go better than four paragraphs of writing ever did.

Templates

Instead of a blank box, you can now start from a format. Poster and Merch are two of the named ones, and there are templates for flyers and product photos.

Pick the format, fill in what you want it to say and the look you want, and you skip the part where you try to describe a poster from scratch.

Comments on the image

Here is the one that fixes my ad problem. You can now place a comment directly on a spot in the image, the same way you would mark up a PDF proof, and tell it what to change right there.

Point at the phone number. Say "this type is too heavy, make it lighter and narrower." Everything else stays where it is.

If you have ever handed a proof back to a designer with a red circle on it, you already know how to use this.

Shareable prompts

When you share an image, you can now include the prompt that made it. The person you send it to can run the same idea with their own photos and their own details.

For a small shop, this is the quiet one. Your designer works out the prompt that produces your house ad style. She shares it. Now the two people covering nights can make one that matches without bothering her.

Five things a newspaper team can do with this next week

1. Spec ads you can actually put in front of somebody. You know the account that has said no for three years. Sketch the layout, describe their business, and walk in with a finished looking ad instead of a rate card. The faster generation means you can make three versions in the parking lot.

2. Fix the one thing that is wrong. Upload the ad or the flyer you already have, comment on the exact spot, and change only that. The August edition of this newsletter was all about the drift problem, where every fix broke something else. This is the feature that addresses it.

3. House ads and subscription promos. Every paper has a hole to fill on page 6 at eleven o'clock at night. Templates plus your own text gets you a clean house ad in about four minutes. Build one you like, save the prompt, reuse it all year.

4. Art for the story that has no art. The opinion column. The tax levy explainer. The brief about a program where no photo exists. An illustration is honest where a fake photo is not, and this is now good enough to be worth using. Read the rules section below first.

5. Event and merch graphics for a community client. The Merch template exists for a reason. The chamber's fun run wants a shirt design. The booster club wants a poster. That is a sellable service you can deliver in an afternoon, and most small papers are already the only design shop in town.

Making the image actually work on your website

This is the part everybody skips, and it is the part that decides whether the image helps your site or drags it down. ChatGPT makes a picture. It does not make a web ready picture. Those are two different jobs.

Four steps, in order.

1 Ask for the shape you need before it draws.

Say it in the prompt. A wide horizontal image for the top of a story. A square for social. A tall one for a phone. Asking up front costs you nothing. Cropping a square into a banner afterward costs you the top of somebody's head.

If the image needs to sit on a colored background on your site, ask for a transparent background while you are at it. Images 2.5 handles those now, and it saves you the cutout work.

2 Shrink the file before it goes up.

What you download is built to look good, not to load fast. Putting that straight onto an article page is how a site gets slow, and a slow site costs you readers and search traffic both.

Run it through a free compressor first. Squoosh, at squoosh.app, is free, made by Google, runs right in your browser, and nothing gets uploaded anywhere. Aim for under 200 KB for a story image. Most images get there with no difference you can see.

Some content systems do this automatically on upload. Check yours once. If it does, skip this step forever.

3 Give the file a real name.

A file named for the date and the tool tells a search engine nothing. A file named windom-eagles-homecoming-parade.jpg tells it everything. Ten seconds of work, and it is the cheapest search traffic you will ever buy.

4 Write real alt text.

Alt text is the written description of an image. Screen readers read it aloud to readers who cannot see the picture, and Google reads it to understand what the image shows. Describe what is actually in it, in one plain sentence.

And here is the part specific to this week: if the image was made by AI, say so in the alt text and in the caption. Not in the fine print. In the caption, where a reader will actually see it.

Here is a prompt that handles the first step and hands you the last two.

I need an illustration for a story on our newspaper website.

The story is about: [one sentence on the topic]

Make it:
- A wide horizontal image, sized for the top of an article page
- An illustration, clearly not a photograph
- Simple enough to read at a small size on a phone

Do not put any text inside the image.

After you make it, give me two things in writing:
1. A one sentence alt text describing what is in the image
2. A suggested file name, all lowercase, words separated by hyphens

Three rules before any of this touches your paper

1. Never generate a news photograph. Ever. No AI made picture of a fire, a crash, a meeting, a game, or a person in your coverage area, no matter how clearly you label it. Illustrations, house ads, promotional art, and client design work are fair game. Documentary images are not. That line is the whole ballgame for a newspaper, and there is no version of crossing it that ends well.

2. Label it where the reader can see it. A caption that reads "illustration generated with AI" costs you nothing and buys you the benefit of the doubt for years. The paper that gets caught not labeling is the one explaining itself at the coffee shop all winter.

3. Do not let it redraw a logo or a trademark. I said this in August and it has not changed. The Realtor logo, the MLS mark, the Equal Housing symbol, a client's own logo. AI gets them subtly wrong every time, and with the real estate marks that is a compliance problem, not just an ugly one. Use the real files.

One more thing worth knowing on the transparency front: OpenAI says it puts invisible watermarking and something called C2PA data into the images it makes. There is a short explainer on that at the bottom of this email.

What it costs

Images 2.5 went out to everybody. All ChatGPT tiers, on desktop, phone, and web, plus ChatGPT Work and Codex. There is no new subscription to buy.

What your plan still decides is how many you get. Free accounts are capped per day, and OpenAI does not publish the number. Paid plans get meaningfully more. If you try this and hit a wall after a few images, that is the cap, not a bug.

If your web developer asks, there are two new versions for building this into software: GPT-Image-2.5 Flare, the fast everyday one, and GPT-Image-2.5 Sunburst, which is slower and more precise for finished campaign work. That is a conversation for whoever handles your tech, not for you.

Try this on Friday

Take one house ad you are tired of looking at.

Open ChatGPT, type @Sketch, and draw the layout you actually want in thirty seconds. Boxes and lines, nothing fancy. Describe your paper's look in a sentence or two. Let it build.

Then find the one thing that is wrong with what comes back, comment on that exact spot, and fix only that.

If it works, save the prompt. That is your template for the rest of the year.

Have a good weekend.