We tried the best image generation model! ChatGPT handles mountains of text and great portraits

Technology companies are constantly competing to create the most capable language model. However, they also attach no less importance to generating multimedia content, especiallyimage creation. While until recently Google could be considered the king with its Nano Banana 2 model, with the Tuesday launch of ChatGPT Images 2.0, OpenAI is taking the lead. We have tested the new model andare bringing you practical examples.

ChatGPT is once again the king of image generation

While a year ago, companies behind AI models primarily enticed with breathtaking images with a wide spectrum of colors, OpenAI, in connection with ChatGPT Images 2.0, focuses primarily on practicality. Not that the model is incapable of generating fascinating sceneries – you’ll see them below, after all – but the priority is precise adherence to instructions, meticulous work with objects, and above all, flawless text in various languages and with high density. The driving force is the ability to think, which the OpenAI model has for the very first time, and also the ability to search for information on the internet.

The creators are simply making it clear that the new model is not just for playing around. It can create top-notch infographics ready for immediate publication on social media, stunning posters, or even magazine pages and captivating comics. It also excels at mimicking handwritten text. In short, it’s no wonder it took first place in the Arena AI ranking with a significant point lead.

Practical examples: poster, recipes, and illustrative image

As soon as I tried ChatGPT Images 2.0 for the first time, I realized how much I regretted that OpenAI hadn’t released the model a few months earlier. It wasn’t that long ago that I had to create a poster with information about the philosopher William James. Since I’m no design expert, and more importantly, I didn’t want to invest time in this specific task, I decided to ask Gemini for help. I attached files with all the necessary information and a portrait of the philosopher and received an image that was so visually appalling I wouldn’t even hang it on a toilet door. That’s why I was even more surprised by ChatGPT’s result – while those more knowledgeable would probably recognize that the poster was created by artificial intelligence, there’s certainly no reason to be ashamed of its final appearance.

A popular example of use when introducing new image generation models is the processing of recipes into a visual form. Therefore, I didn’t hesitate and inserted instructions for preparing Raffaello balls copied from the internet into ChatGPT. I dare say the result is fascinating – similar to the philosopher, it’s remarkable how much text the AI handled. I also highly appreciate the resulting design, which I would immediately use to enrich my article as a cooking blogger (if I were one, of course).

Since I liked this design, I decided to use ChatGPT Images 2.0 to create more recipes for my favorite treats – potato salad and strawberry milkshake. The consistency is great, and I like how the AI adapts the color palette to the given dish.

Finally, I needed an illustrative image for an article about unauthorized access to the Anthropic Mythos model (which was later published on the website on Thursday afternoon). I was somewhat inspired by Reuters, where they like to use a motif with a smartphone and a company/service logo. So, I attached the Anthropic logo to ChatGPT and entered the prompt:

„generate an engaging widescreen image featuring a phone with the Anthropic logo against a background with the head of a humanoid artificial intelligence. the image should be widescreen and colorfully attractive“

However, I didn’t like the aspect ratio of the first version of the image – I knew that after cropping, it wouldn’t fit into the grid I had to follow. So, I clicked on the image and selected a new 16:9 aspect ratio from the menu. After a moment of generation, ChatGPT came up with a new image.

Of course, you might have noticed that the prompts I used were quite trivial. Artificial intelligence experts spend tens of minutes, sometimes even hours, crafting prompts to get the most precise results possible. In this way, I rather demonstrated how even a complete layman, who simply doesn’t want to create lengthy detailed instructions, can extract quality results from ChatGPT.

I can imagine that the infographic above regarding the Redmi Watch 6 would be more perfect if I had played around with the prompt more. But considering I spent about three minutes on it in total, including the waiting time for its creation, it’s quite decent.

It will edit your photos exactly as you wish

ChatGPT Images 2.0 can also handle changing photo styles excellently. My selfie from Spain, after entering the prompt „Turn this photo into an animated avatar. Keep all details, including the background, just redraw it in an animated form.“ looks like this:

Equally interesting was the conversion of a photo into a sci-fi image full of neon lights with the trivial prompt „Rework this image into a cyberpunk style with neon lights“

Last but not least, I remembered that after a long time, I could change my profile picture as an editor or on platforms like LinkedIn. So, I attached my three selfies to ChatGPT and the prompt „Create a studio portrait for me as a new profile picture for my editor’s profile. I’m attaching my photos so you know what I look like. I will be wearing a black shirt with the top button unbuttoned and a serious, yet smiling expression.“.

Although I am satisfied with the result in terms of accuracy and adherence to instructions, I will still have to decide whether this type of photograph truly suits me enough to present myself publicly with it…

Gallery of examples: portrait, Android, and more

What do you think of the new image creation model?