Why is AI so bad at generating cross-stitch alphabet patterns?

My first attempt at needlecraft was in January 2024. It was an embroidery kit that I never actually finished, but it’s still framed and hanging on a wall in my room. Shoutout to the embroiderers of the past because it was honestly so painful to repeatedly push a needle through fabric.

A couple months later, after my fingers had recovered and my motivation had returned, I gave needlepoint a go. It was significantly less painful; thanks to the open weave canvas the stencil was printed on. It’s a great, crafty activity because I could do it while watching something (and therefore feel less guilty about binging entire shows). And I also loved gifting these handmade projects to people.

Earlier this year, I decided to give cross-stitching a proper go. It’s similar to the type of needlepoint I had already done, where I was stitching in one diagonal (a forward-slash); and cross-stitching just meant that I’d add another stitch over this one (a backward-slash) to then create an ‘x’. I wanted to create my own cross-stitch designs so that I could make personalised gifts. I got plenty of inspiration from Pinterest for ornamentation. But I also wanted to include a meaningful word in the design.

I was on the hunt for a simple, aesthetic, free, block-style font pattern. And it was a struggle to find something I liked. So, I did what anyone my age does, and asked ChatGPT.

This is what it gave me:

I even tried to ‘guide it’ further by prompting it to ‘think’ of cross stitch alphabets as pixel art, but this is what I got in return:

It was interesting – the LLM must really like the letter B. But the takeaway here is that ChatGPT was clearly horrible at this task. Why?

(Note: This all happened months ago, but I'm only researching this on a deeper level now.)

I was curious to see if this issue persisted across different models. So I gave ChatGPT, Adobe Firefly, Canva AI, DeepAI and Google Gemini the exact same prompt (on 5 Aug 2025). Here are the results:

Prompt 1: "Generate an image of a lower-case cross-stitch alphabet design, in a block font"

These are my observations (no disrespect to these genAI’s, they are trying their best, I guess):

• ChatGPT thinks there are 25 letters in the alphabet; but has the grid-style needed for cross-stitch.
• Adobe Firefly has great taste in colour, but the output is not cross-stitch compatible.
• Canva AI is a big miss… and also mysteriously obsessed with the letter B.
• DeepAI seems to have poorly split the alphabet in two.
• Google Gemini has all the letters, but appears to stutter by repeating a couple.

I then wanted to see if this result was because my prompt was simple and lacking the detail of what I was looking for. So I tried something more detailed:

Prompt 2: “Generate a complete cross-stitch, lower-case alphabet design from ‘a’ to ‘z’ in a bold, block-style font. Each letter is stitched using X-shaped cross-stitches and arranged clearly in rows on a white square grid background, resembling Aida cloth. The grid should be evenly spaced and visible, with each square representing one stitch. The letters should be formed cleanly with black thread for high contrast. The overall aesthetic should be neat, simple, and suitable for beginner cross-stitchers. Include consistent spacing between letters and rows, and ensure the entire alphabet is visible in the image. No decorative embellishments, just the alphabet in clean, block-style cross-stitch on a grid.”

Observations:

• ChatGPT is so close, but it doesn’t output all 26 letters and is honestly not very aesthetic.
• Adobe Firefly is clearly tripping, and it ignored like 97% of the prompt.
• Canva AI is also spitting garbage.
• DeepAI – no one asked for numbers…
• Google Gemini – correct number of letters! but clearly not cross-stitch friendly.

Given the above, it’s safe to say that AI is bad at generating cross stitch alphabet patterns. Why?

AI image generation is deeply rooted in maths, physics and the concept of diffusion. Just like how tea clouds up upon the addition of milk, AI models start with a ‘blank/random’ slate and then use maths to guide a creation from text.

A transformer is what works on creating an image from this randomness. And no, this kind of transformer is nothing like the cool trucks from the movie series; they simply work step-by-step on changing a text input sequence into an output (an image in this case), via a ‘denoising’ process.

AI doesn’t encode words, or even letters, the way that humans do. But just like a misguided toddler, AI can form skewed models based on what data is presented to ‘em.

The 5 AI models I tested above are trained on huge datasets, featuring all kinds of images. They essentially learn their own patterns (in terms of statistical regularities) between text and images. These models are great at generating continuous patterns, such as landscapes, because the transformers iterate over many steps to create something true to the shapes, colours, and textures it is trained on.

The distinction here is that cross-stitch is discretised; existing entirely on a binarised grid structure. The continuous diffusion models that AI run, paired with the fact that transformers cannot enforce strict spatial constraints, means that it can’t replicate letters with the symbolic precision that the trained human eye can accept as accurate.

Plus, AI can't judge aesthetics the way we do. AI has no concept of beauty; only some understanding of (or hallucination) from patterns that it is fed. This is good news for the dreamers & creatives out there – keep working your magic because your friendly neighbourhood LLM’s cannot create like you.

Here's my (self-made) cross stitch alphabet:

And here's my first ever cross-stitch creation (from March 2025)!! It was a graduation gift for my brother and a reminder to take care, always.

PS. I could not have made this post without the work of others, so shoutout to:

This fabulous video explaining the physics/maths of AI image generation.

Mama Geek for the free Cross stitch grid Procreate templates

And Yarn Tree for the font inspiration, with their free Times Roman Alphabet Chart.