Can ChatGPT generate images?

The short answer

Yes. ChatGPT can generate images from a text description, and it can usually refine them through conversation, which is its genuine advantage: you can say "same thing, but wider and less blue" instead of rewriting a prompt.

Whether it is the right tool depends on the job. For a quick illustration or a concept sketch, it is excellent and convenient. For consistent style across a set of images, precise aspect ratios or production work, dedicated tools do better.

Availability and limits vary by plan, so check OpenAI's pricing page for what your account includes.

What it does well

• Concepts and mockups. A rough visual to explain an idea in a meeting.
• Iteration through conversation. Adjusting in plain language rather than rewriting parameters.
• Simple illustrations for slides, internal documents and social posts.
• Understanding complicated instructions. It parses a long, awkwardly worded description better than most dedicated tools, because that is what it was built for.

Where dedicated tools win

Midjourney for aesthetic quality and stylistic control. Its parameters and style handling give more repeatable results, and there is a guide to writing prompts for it.

Adobe Firefly when commercial licensing and integration with Photoshop and Illustrator matter. Adobe trained it with commercial use in mind, which is why design teams reach for it. Its product page sets out the current terms.

Leonardo AI for fine control, model choice and iterating on a fixed character or style.

Anything needing text in the image. All of them struggle, and a design tool with real typography is still the answer for anything with words on it.

How to prompt for images

Different discipline from text prompting. Image models respond to description rather than instruction, so front-load the concrete nouns.

A workable structure:

Subject, then setting, then style, then lighting, then framing.

"A ceramic coffee cup on a linen tablecloth, morning light through a window, soft shadows, shallow depth of field, photographed from slightly above, muted warm palette."

Things that improve results:

• Be concrete. "Cozy" is vague. "Warm lamp light, wooden surfaces, slightly cluttered" is not.
• Name the medium. Photograph, watercolor, 3D render, line drawing, oil painting.
• Describe the light. It changes the mood more than any other single word.
• Say the framing. Close-up, wide shot, top-down, eye level.
• Say what to leave out. Some tools support negatives explicitly, and all of them respond to "no text, no people".

The two things to check every time

Hands and text. They have improved and they are still where errors appear first. Zoom in before you publish anything.

Rights. Whether you can use an AI image commercially depends on the tool's terms and your jurisdiction, and copyright over generated images is genuinely unsettled in many countries. If it is going on a client's packaging, read the terms rather than assuming. This is the most common way people get themselves into trouble with these tools, and it is entirely avoidable.

When not to use AI images at all

• When a real photo exists. A picture of your actual product beats a generated approximation of it.
• When brand consistency matters. Getting the same character or style across twenty images is still hard.
• When the client has not agreed to it. Ask first. Some clients have policies, and finding out afterwards is worse.
• When it is a diagram. A chart or a diagram should be drawn, not hallucinated.

Where to learn the craft

Getting good at AI images is a real skill: prompt structure, iteration, and knowing which tool suits which job. The AI & Automation school covers Midjourney, Leonardo, Firefly and Runway for video.

One subscription opens all of them, which is the sensible way to work out which tool fits your work without buying four of them.

Generate in whatever is convenient. Produce in whatever gives you control.