ChatGPT Images 2.0 is genuinely fantastic
How to use the most powerful image generation AI model in the world.
As most of you are aware by now, 100% of my LLM usage for knowledge work is done through Claude. That’s why I’ve written so many guides about Anthropic’s incredible announcements recently, including Cowork, Projects, Skills, and Interactive Visuals.
But when it comes to image generation, it’s a completely different story.
My go-to model for the last two months has been Google’s AI Studio using Nano Banana 2 (NB2). The reason for this has been the fine-grained control over initial output quality (512px, 1K, 2K, or 4K) and aspect ratios (1:1, 16:9, 9:16, 3:4, 4:5, 4:1, etc).
Yet what has always been lacking is fine-grained control over iterations.
You see, when you generate an initial design from Google’s AI Studio, it nails the quality + dimensions. But if you wanted to specify a certain area to change in your output, you’d have to describe the change using natural language in your prompt. This turns out to be a real pain, as you write things like “change the wording of the title in the upper-left box to…” only for it to fall flat on its face, for you to leave in frustration, and to then have to reach for Figma or Canva to attempt a botched design job yourself.
The only alternative was to generate the output in the Gemini app itself. However, that comes at the cost of losing aspect-ratio control alongside having to put up with the Gemini watermark in the bottom-right corner of your output.
But it seems as though there’s been a shift in the air.
Nano Banana 2 vs GPT Image 2
OpenAI released ChatGPT Images 2.0 earlier this week, and this thing is genuinely fantastic.
The GPT Image 2 model is now my default for ALL image generation tasks. It trumps AI Studio’s setup with NB2 because of two things. Well, actually three. Firstly, superior initial generations. It tightly understands your prompts, and the vibrancy of the initial output is remarkable.
Secondly, iteration. You can now quickly iterate on designs with the “Select” tool by lassoing items you want to edit and use the “Aspect ratio” control tool to specify output dimensions (see below). These appear once you click on any given output generated by ChatGPT Images 2.0.
Thirdly, and most importantly, the outputs don’t look AI-generated. If you create something using NB2, you know it was created using NB2. There are certain tells throughout the typography, icon selection, and colouring that are immediately visible. Whereas if you create something using GPT Image 2, you don’t have a clue. Take a look at these three examples below:
Prompt 1: “A slide about renewable energy”
Nano Banana 2:
GPT Image 2:
Prompt 2: “An infographic about Tesla”
Nano Banana 2:
GPT Image 2:
Prompt 3: “A bar chart showing global population growth over time”
Nano Banana 2:
GPT Image 2:
The difference is night and day.
I find NB2’s vibe almost has a “cartoonist” default, with a single font and consistently chosen borders, shadows, and icons that don’t really change from output to output. GPT Image 2, on the other hand, looks like it was created by a “professional.” The outputs are clean, not “overdesigned”, and have a taste that is currently unrivalled. What’s more, the number of hallucinations in GPT Image 2 I’ve found to be significantly lower than in NB2, just take a look at the outputs from prompts 2 and 3 above.
In prompt 2, “An infographic about Tesla”, NB2 falls short when it comes to the details of smaller components that make up the composition of a larger output. Particularly section 5 “Impact & Scale”, where the charts are garbled, and section 6 “The Future”, where “terrawatt scale production” is repeated twice. It’s these nitpicks that can be incredibly frustrating when you can’t quickly iterate and iron them out with an easy-select tool.
In prompt 3, “A bar chart showing global population growth over time”, NB2 decided to place the Y axis just before the last projected bar for the year 2100. It also colours the tops of the projected bars without a clear key to what the colours actually mean. GPT Image 2 just keeps it simple. Exactly what I asked for, and it doesn’t feel the need to overcomplicate things with backgrounds or make it overly annotated.
These were all one-shot attempts. I took the first output from each model and compared them head-on.
Something to pick up on, and a key difference between NB2 and GPT Image 2 is speed.
This is where NB2, by contrast, blows GPT Image 2 out of the park. For Google’s model, I’ve found that each generation takes anywhere from 20-25 seconds. For OpenAI’s, each generation can span from 40 seconds to well over a minute. GPT Image 2 is >2x slower than Nano Banana, which is significant but not significant enough to justify the trade-off in output quality. I’d happily wait an additional 30 seconds to receive a superior output. Only if this were a couple of orders of magnitude higher (20 minutes), then I might reconsider. However, the opportunity cost remains low. The alternative is settling for a sub-par output from a now-second-best model, paying a human designer to produce it, or spending your own time making it yourself—all of which cost considerably more than 30 seconds of waiting.
A useful workflow
Now, changing tack slightly, and after experimenting with GPT Image 2 for a while, I thought I’d walk you through a step-by-step workflow that really stretches the legs of ChatGPT’s new image model and showcases its true capability. This is how you make it maximally useful, and for today, we’ll be using turtles…