I tried ChatGPT’s new image generator, and it shattered my expectations

OpenAI may have kicked off the text-to-image generation craze with its DALL-E model, but since those earlier glory days, the AI company’s offering has been lapped by much more capable image models. As a result, when OpenAI released its latest and greatest GPT-4o image generation model, I was skeptical. After testing it, I have changed my mind entirely. 

Getting started

When DALL-E first launched, it lived on its standalone website; since then, it has moved to ChatGPT. The move came with many benefits, including being able to ask the AI chatbot for an image you want in the same interface where you’re already chatting about something else, thereby eliminating the need for constant context switching.

With the release of GPT-4o image generation, OpenAI kept this convenient format, switching the default image generator from DALL-E to GPT-4o for paid subscribers. As a result, it was super easy to start creating new images from my ChatGPT Plus account. All I had to do was enter the prompt for what I wanted to see, and then it would generate them. Users can also access it from the Sora interface. 

Also: How to use OpenAI’s Sora to create stunning AI-generated videos

Beware: You can still generate images similarly if you are a free user. However, if you’re unimpressed, that’s because even though at launch, the model was announced to be coming to all users, including free ones, OpenAI CEO Sam Altman announced a day later that the rollout to the free tier would now be “delayed for awhile.”

The images 

The moment you have been waiting for — the images. After you insert a prompt, the AI outputs the generation in under a minute. The process does take a bit longer than it used to, but the images are worth the wait, delivering lots of details, texture, realism, and even text accuracy. Instead of describing it, I will include examples below so you can see for yourself. 

Prompt: Can you generate a realistic image of a chameleon, up close, shot as if it were in National Geographic in 16:9 ratio?

Prompt: Can you generate an image of a laptop open on a desk that says, “This model is so good that it can even get text and hands right, which are usually major challenges for AI models,” with hands typing on a keyboard in 16:9 ratio?

Prompt: Can you generate a realistic photo of a close-up of a woman in a crowd in Times Square looking at the camera and smiling, with the quality of one taken on a DSLR?

As seen above, the image generator does a great job of adhering to the prompt and delivering high-quality, realistic images. However, when testing an AI model, one of the true performance metrics is how it compares to competitors on the market. To give you a good indicator of this, I made it generate the same prompt I tested across all of the major AI image generators, including Midjourney, Google’s Imagen 3, Adobe Firefly, and more. 

I am attaching GPT-4o’s rendition below. You can see how it fares against all of the other AI image generators in this article, including DALL-E’s rendition, which clearly is far behind what the new model can do. 

Prompt: Can you generate an image of a vibrant, realistic hummingbird perched on a tree?

Other notable features

Even though the quality of the images is perhaps one of the model’s biggest wins, there are other benefits as well. One of the biggest is that it lives in the chatbot’s interface, which makes it easy to tweak the generations with simple natural language prompts. Also, because the chatbot has the context of what you just asked it, it can consider that in building the image. 

For example, if you are chatting with it about throwing a birthday party, you may be able to say, “Can you now create an invite that has the information above on it?” instead of having to retype. For example, I started chatting with ChatGPT about throwing a housewarming, and when asking to make it create an invite, I didn’t have to repeat the information I previously said.  

You can also upload reference images and then ask ChatGPT to create a different version or use them as elements of a new one. For example, you can input it as a selfie and have it generated in anime style, as seen in Altman’s new X post. 

All of these customization features make it a really strong offering for creatives, who can also request that it be rendered on a transparent background or incorporate brand style guides such as hex codes or logos. 

Speaking of Altman, I was able to generate an image of him wearing a party hat. I could do so because the new model has much looser safeguards, meant to allow users to lean into their creative freedom. The blog post announcing the model noted that it limits what can be created when real people are in the context, including “particularly robust safeguards around nudity and graphic violence.” 

I can’t tell if there is a practical use case for this feature, but it is a notable change I needed to try out for myself. When I tried to create an image of Mickey Mouse, it said it couldn’t due to copyright implications, so it seems not all public figures are fair game. 

Overall

Overall, the GPT-4o image generator is a big win over the DALL-E models and perhaps among the best of the many I’ve tested. Is it worth the $20 per month? If you are just interested in high-quality image generation, there are still free versions you can explore that are really capable, such as Adobe Firefly or Google’s Imagen 3. 

Also: The best AI image generators: Tested and reviewed

Having said this, if you are a frequent ChatGPT user, the upgrade to ChatGPT Plus gets significantly more enticing. With this upgrade, you will have access to all of OpenAI’s latest and greatest chatbot features, as well as high-quality image and video generation, all for $20 a month, which is not a bad deal, especially considering other offerings on the market. For example, Midjourney’s subscription starts at $10 per month and only offers image generation. 

Want more stories about AI? Sign up for Innovation, our weekly newsletter.