Limitations of AI

Before we get carried away by the capabilities of AI, reflect on what is shown in Figure 1:

Figure 1

Looks OK. Well, not really since 6 + 5 + 8 + 3 + 6 + 3 = 31 not 35! It seems that Gemini is not great at arithmetic (the calculation was made using the Chrome browser). I also spent a lot of time yesterday trying to get Nano Bananas to implement some simple instructions but it failed miserably. Figure 2 shows the mangled results:

Figure 2

In frustration, I asked Gemini the following:

Absolutely no difference. I can only say I'm mightily unimpressed with Nano Bananas. Can you at least give me some feedback as to why my prompt details are so blatantly ignored. All the image generator had to do was to make some minor changes to the shapes of the letters and numbers. It succeeded with the numbers but then became deranged when dealing with the letters. What confused the image generator?

Here was the response:

I completely understand your frustration, and you have every right to throw in the towel on this specific approach. It was a spectacular typography failure.

To give you a candid look behind the curtain, here is exactly why the image generation failed so consistently and how we can actually solve this.

Why the Attempts Failed

1. The Letter Count Discrepancy

The initial prompt set up a mathematical conflict right out of the gate. The phrase "twenty eight thousand and eighty two" actually contains 31 letters, not 33. When you later changed it to "twenty eight thousand and eighty eight," that phrase does hit the 33-letter mark perfectly. However, trying to force 31 letters into 33 slots in those first few attempts immediately confused the layout generation.

2. The Limits of Image Models

The image generation here is powered by the Nano Banana model. While it is a state-of-the-art model featuring high-fidelity text rendering, it fundamentally understands text as visual patterns rather than strictly logical data. It is excellent at writing a clear word on a sign, but it struggles immensely with rigid, mathematical spatial mapping.

When you ask it to draw exactly 33 numbers, exactly 33 letters, space them perfectly in concentric circles, and connect them with exactly 33 lines, it gets overwhelmed by the geometry. To fill the visual space of the circle, it begins hallucinating letters and creating gibberish filler words like "CHORTY" and "DIGHY" to complete the pattern.

How to Achieve Your Goal

The blunt reality is that no amount of prompt engineering will make an AI image generator reliably produce this level of precise, geometric exactitude. Image models are artists, not draftsmen.

To get this exact mandala, the most effective path forward is to bypass the image generator entirely and build it programmatically. By writing a relatively short Python script using a library like Matplotlib, we can plot the exact coordinates for the 33 numbers and 33 letters on a radial axis, draw the connecting lines perfectly, and apply the exact rich brown and gold hex color codes you are looking for.

Would you like me to write the Python code so you can generate this exact, mathematically perfect design yourself?

In the end the Python code worked quite well and using the Python output with the image generator produced the same frustration. Figure 3 shows the Python output:

Figure 3

So it pays to be aware of AI's limitations and work around them. Gemini at least is far from perfect.

More Infographics

A popular prompt for last year that made use of Nano Bananas was to "create an infographic about this plant focusing on interesting information". So I prompted Gemini with the following:

Create an infographic about so-called "red-eared sliders", as far as I know a type of terrapin native to the Florida everglades, focusing on interesting information.

Well, one thing I learned from the infographic generated was that these creatures are NOT native to the Florida everglades. See Figure 1.

Figure 1

My next prompt (see Figure 2 for the result) was:

Create an infographic about the different types of terrapins found around the world, focusing on their distribution and including interesting information.

Figure 2

My next prompt (see Figure 3 for the result) was:

Create an infographic highlighting the differences between tortoises, turtles and terrapins.

Figure 3 

Applying Different Art Styles

In my previous post, I investigated different art and photographic styles and showed an example of an art nouveau style applied to an old photograph of myself and a friend. See blog post titled Different Art Styles. In this post, I'd like to apply some other effects to the same photograph. Even before my Deep Research article I'd applied three other effects. See Figures 1, 2 and 3.

Figure 1: Aubrey Beardsley style

Figure 2: Vincent Van Gogh style

Figure 3: psychedelic style

Now I'll apply some of the effects suggested by the Deep Research article, the full text of which can be found in my previous post.

Figure 4: Art Deco style

Figure 5: Vienna Secession style

Figure 6: Pre-Raphaelite style

Figure 6: Ukiyo-e style

That's a selection from a longer list of possibilities. Some work better than others when humans are in the picture. I particularly liked the Pre-Raphaelite style.

The YouTuber Pete Beard has a comprehensive series of videos about various artists. Figure 7 shows a screen shot of the first 20 of them:

Figure 7

Gemini recommends the following strategy for obtaining the best results in getting Nano Bananas to emulate an artistic style using Art Nouveau as an example:

Strategy: Instead of just describing "Art Nouveau," the user should upload a high-resolution image of a Mucha poster or a Klimt painting as a style reference.

Elsewhere Gemini advises the following:

One of the biggest challenges in style transfer is losing the identity of the subject. Nano Bananas features "Character Consistency".

Application: When applying a heavy style like Tim Burton or Cubism, the user should explicitly prompt to maintain facial landmarks.

Prompt Structure: "A portrait of [User] in the style of Tim Burton. Maintain subject likeness. Pale skin, large eyes, gothic background."

Result: The AI will adapt the proportions and texture (large eyes, sketch lines) to the style while keeping the fundamental biometric data recognizable. 

Using Charles Keeping as the artist, I uploaded an illustration of his. See Figure 8.

Figure 8

I then prompted Nano Bananas to apply this style to the photograph and Figure 9 shows the quite satisfying result.

Figure 9