6.0/ 10
SCORE

ChatGPT AI Image Generator

Pros

• Available with free and paid ChatGPT plans
• Decent prompt adherence
• Versatile styles

Cons

• Long generation times
• Typo-ridden text
• No advanced editing tools

ChatGPT is practically synonymous with generative AI, so it wasn’t too surprising when its parent company, OpenAI, announced earlier this year that it’s getting into the AI image business. With GPT-4o’s new text-to-image capabilities, the internet exploded with ChatGPT’s Studio Ghibli-esque images. As CNET’s resident AI image and video reviewer, I was particularly excited by the company’s promise that ChatGPT could generate images with a lot of text, something that other AI services typically can’t do. But my wild ride testing ChatGPT’s image generation prowess proved one thing: ChatGPT might be a power player as an AI chatbot, but it’s an amateur image generator.

I spent many hours with ChatGPT, generating close to 100 images using several of OpenAI’s models, including 4o, o3 and o4-mini. There were some recognizable similarities between ChatGPT and OpenAI’s other image generation program, Dall-E. They both have the same conversational editing flow, which is one reason why Dall-E 3 was a CNET Editor’s Choice award recipient last year. But there are a lot of differences, and they aren’t all for the best. 

ChatGPT is slower than any other program I’ve tested, much slower than Dall-E. Unlike Canva, Midjourney and Leonardo.Ai, ChatGPT only generates one image per prompt, which doesn’t give you a lot of variety or choices. Most distressingly, it doesn’t have advanced editing capabilities. And as you’ll see in the images I created, that’s going to be annoying when you want to make text-heavy images. The most comparable program is Meta AI, which is also primarily a chatbot. Between the two, I would give the edge to ChatGPT for more creative and versatile results.

If you’re used to using ChatGPT, it will feel familiar. But if you’ve used a lot of other graphic design and editing programs, or even just other AIs, you’re going to be disappointed. OpenAI’s image generation is best for users looking for a quick image tool or those who are less creatively inclined but still want to bring their ideas to life without much heavy lifting. It is certainly not for professional creators looking to expand their digital toolkits.

Here’s how ChatGPT fared against my testing, including how good the images were, how well they matched my prompts, how you can fine-tune results and how quickly they arrive.

(Disclosure: Ziff Davis, CNET’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)

How CNET tests AI image generators

CNET takes a practical approach to reviewing AI image generators. Our goal is to determine how good it is relative to the competition and which purposes it serves best. To do that, we write AI prompts based on real-world cases, such as rendering in a particular style, combining elements into a single image and handling lengthier descriptions. We score the image generators on a 10-point scale that considers factors such as how well images match prompts, creativity of results and response speed. See how we test AI for more.

You can create images using ChatGPT with any plan. You’ll have limited access with the free plan. Upgrading to the Plus ($20 per month) and Pro ($200 per month) plans gets you higher rate limits, along with other ChatGPT features. Even on the paid plans, you may run into wait times, where OpenAI’s servers are overloaded and your requests are delayed.

OpenAI’s general privacy policy, which the company said includes image generation, says it can train on your personal information and content to improve its services, including using your prompts and any images or files you upload. To opt out, go to Settings > Data control and turn off Improve the model for everyone. Abusive, illegal and other malicious uses of ChatGPT’s AI image generator are not allowed and should be prevented, according to GPT-4o’s system card. You can read more in OpenAI’s safety policies.

How good are the images, and how well do they match prompts?

ChatGPT is pretty versatile when it comes to style. It handles photorealistic stock imagery, as well as more whimsical, cartoonish looks. The images had a fun, engaging style. ChatGPT images do not have a visible watermark denoting they’re AI-generated, so always be sure to disclose that when you share them. Images do have C2PA metadata that indicate they were AI-generated.

The main differentiator between ChatGPT and other AI image services is that ChatGPT can generate text-heavy images. Most AI image generators can’t generate more than a few letters; ChatGPT is definitely the most impressive AI image generator I’ve seen when it comes to creating legible, English-language text.

Of course, being able to create letters is one thing, but putting them in a meaningful construction is another. This is where ChatGPT struggles. I tried to generate a lot of education posters — topics that require some descriptions but also allow the generator to show off its creativity. Some turned out okay. Others were a complete disaster.

ChatGPT does best when you give the chatbot a website or text passage to reference when creating text-heavy images. Even then, it’s not always great.

And I really have no idea what happened here. My apologies to Jane Austen and the cast of the 2005 movie adaption of Pride and Prejudice.

Despite OpenAI saying its models should not be able to reproduce copyrighted or protected imagery, I was able to sometimes. ChatGPT created a fake product stock photo of a Hydroflask water bottle with the real logo, even though the chatbot itself told me it shouldn’t be able to do that. I also tried to get it to recreate the Teenage Mutant Ninja Turtles in a few photos, and while prompts where I asked for them by name were blocked, other prompts describing them as “human-like turtles with colorful masks” produced very similar results. This is likely a result of OpenAI’s broad training data. (Side note: OpenAI is being sued by a number of companies and artists alleging copyright infringement and misuse of their protected content.)

Other image generators I’ve tried haven’t been able to replicate brand logos like ChatGPT. I strongly advise against creating AI images that could infringe on protected content, especially with name brands or a person’s likeness.

I primarily tested ChatGPT on my laptop through the website, but I did create a few images using the iPhone app. The experiences were comparable, and it is nice to have the option to create on your phone if that’s your preference.

ChatGPT has pretty good prompt adherence for your initial prompt, which means that it can follow directions. Like with any image generator, practicing good prompt writing will help you achieve your artistic vision more easily. That means including a lot of detail — the aesthetic, style and general vibe you want, in addition to the characters and setting. ChatGPT has a harder time when you’re editing photos after, so I recommend really taking some time to craft your perfect first prompt. That will help you get usable final results without too much fuss.

How engaging are the images?

The images I created with ChatGPT were pretty engaging, very colorful and bold. I enjoyed how creative the model was, and although it’s not a scientific or quantitative measurement, I felt like ChatGPT could create an interesting image with nearly any kind of prompt. They weren’t the most intricate or detailed images — I would give that title to Midjourney or Leonardo.Ai — but they were decent. If you put in your prompt that you want a lot of fine details, you’re likely to get better results.

Can you fine-tune results?

Yes, you can edit images after you’ve generated them. The main way you do this is by sending a follow-up prompt asking for changes. You can highlight the specific area you want adjusted for more accurate adjustment. Click on the image to enlarge it and tap the paint brush icon in the upper right corner to highlight the specific region where you want changes to occur. You can then tell ChatGPT what you want fixed via a text prompt. I recommend making these region selections every time because you’re more likely, though not guaranteed, to get your desired changes where you want them. 

If you’re generating a text-heavy image, you may be able to get the service to fix typos. I tried a bunch of different methods, including highlighting the region where the typo is, rerunning the prompt and just straight up telling ChatGPT it messed up. Sometimes it will fix them, but you are absolutely going to need to be patient and work through those errors one by one. In my experience, the process was like fighting a hydra of errors —  fixing one error often spawned three new ones. It was a frustrating, usually unsuccessful process.

I also ran into an issue where ChatGPT would repeatedly cut off the far left side of the image, meaning I had to go back in and ask for it to extend the left side so the whole image would be visible. While it was an easy fix, a simple aspect ratio issue is something that I expected a company like OpenAI would be better at addressing. Regenerating to get the whole image visible also opened the door for the program to change other, nonselected areas, which inevitably led to more errors.

It’s easy but annoying edits like these that make me wish OpenAI would integrate advanced editing tools like it has in Sora, its AI video generator. I liked the conversational flow for brainstorming and making big changes, but there were many times I was frustrated with the minimalist approach to editing features. It’s super disappointing, and I hope OpenAI brings in better options in a future update.

How fast do images arrive?

ChatGPT has the longest generation time of any AI image service I’ve tested. It takes anywhere between 1 to 2 minutes for an image to arrive, even using its most recent and supposedly fastest model, GPT-o4 mini. Most AI image generators take 20 to 40 seconds for a batch of four images. ChatGPT takes more than twice as long to create a single image. 

Watching the image load on my screen, from the top down, is reminiscent of watching my parents’ old printer try to eke out a single paper — agonizingly slow and frustrating. For people like me who are so used to near-instant responses, the speed is exasperating. I have to believe that generating images with a lot of text — forcing the program to use a lot of energy — is part of the reason for the delay.

ChatGPT is a chatbot first, an image tool second

After all my testing, I was left feeling disappointed. When a big tech company like OpenAI invests its resources and talent into a new feature like image generation, I expect something that can either match or surpass its competition. But ChatGPT is first and foremost a chatbot, and it showed throughout my testing. The primary purpose of a text-extruding machine like ChatGPT is to generate text. But like how ChatGPT’s regular text results can be filled with made-up information, its AI images with text were wonky and typo-riddled.

The lack of advanced editing tools is a real bummer, because any way to remediate the typos and errors would’ve made this program a lot more usable. Instead, it was a frustrating experience to have to fix individual things via prompting. More often than not, fixing one error would inadvertently create new ones in the subsequent generation. The conversational flow is nice, but it’s not a viable substitute for specific post-generation editing tools. I wouldn’t mind it’s slow pace and ability to only generate one image per prompt if that image were good and easily editable. 

It’s difficult for me to believe that this is the same company behind Sora, OpenAI’s video generator, which has a much smoother, professional experience. I’m hopeful that the company might eventually bring these tools over to ChatGPT. Otherwise, I’ll stick with using ChatGPT only for the occasional search query.

For more, check out our guide to writing AI image prompts and best AI chatbots.