Can Claude and ChatGPT generate images for ads?

Not on its own. As of August 2026 there is no diffusion model behind Claude, no in-house equivalent of DALL·E or Imagen, and Anthropic reaffirmed that position in April 2026. The Claude 5 family did not change it. If you have been typing image prompts at Claude and getting text back, nothing is wrong with your account.

What Claude does with images is read them. It analyzes what you upload, writes SVG and code-drawn graphics in Artifacts, and through Claude Design builds layouts and decks out of code rather than photographs. That is a genuinely different capability, and for diagramming or interface work it is often the better one.

The important exception: Claude can connect to an outside image generator through MCP, the Model Context Protocol. Once a generator is exposed as a tool, Claude calls it like any other tool and images come back into the conversation. So the honest answer to “can Claude generate images” is no by itself, yes with a connector.

Yes, and the raw image quality is genuinely good. Ask for a product on a marble counter in soft morning light and you will get something that looks like a photograph. For moodboards, concept exploration, and social posts where the product is incidental, this is often all you need and it costs nothing extra.

Ad creative asks for something different, and that is where the distance shows up. An ad is not one nice image. It is your actual product, rendered accurately, with legible copy positioned so it does not cover the thing you are selling, in several sizes, in several versions you can test against each other. Every one of those requirements is a place general-purpose generation gets uneven.

Four places, in roughly the order they will annoy you.

Text baked into the pixels. Image models have got much better at spelling, but the text becomes part of the picture. Changing “20% off” to “25% off” means regenerating the whole image and accepting a different photo. Over a week of testing headlines, that is the difference between editing and starting over.

Product fidelity. This is the expensive one. Give a general model a photo of your bottle and ask for it on a beach, and it will produce a bottle that resembles yours. The label text goes soft or invented, the cap proportion shifts, the brand green drifts a shade. Fine for a pitch deck. Not fine when a customer receives a product that does not match the ad they clicked.

Consistency across a set. Ads are judged as a campaign. Ten images from ten prompts look like ten different brands, because nothing carries your palette, type, and framing from one generation to the next.

Nothing is editable afterward. You get a flat file. No layers, no way to nudge a headline off the product, no way to export the same creative at 4:5 and 9:16 without regenerating both and hoping they match.

Often, and it is worth saying plainly. If you are exploring a concept before committing budget, generating backgrounds and textures with no product in them, making internal decks, or producing images where nothing has to match a real object, a general model is faster and cheaper than anything purpose-built.

It is also the right tool when the constraint is your own taste rather than the software. If you do not yet know what you want the ad to look like, no amount of tooling fixes that. Look at what is already working in your category first.

Two changes fix most of it, and you can apply them with any tool. Keep the text out of the generated pixels and place it as a real layer on top, so copy stays editable and can be moved off the product. And verify the product rather than trusting it: put the generated image next to your original photo and look at the label before you spend money on the ad.

That second step is the one people skip, and it is the one that catches the failure that actually costs you: a beautiful ad for a product that is subtly not yours.

Dovey is built around exactly those two rules. Your product goes in as a reference and comes out placed as photographed, with every image checked against your original so a drifted label gets regenerated before you see it. Text stays a live layer, so rewriting a headline or moving it costs nothing and changes nothing else. And because Dovey runs as an MCP server, it is also a straightforward answer to the first question in this article: connect it and Claude can generate real ad creatives through it, with the same capabilities the web app has. It is $1 for three days, which is enough to run your own product through it and compare against whatever you are using now.