An accidental comparison of AI models generating pictograms

Recently I wanted AI to generate a small-sized pictogram for a small stamp, maybe 20mm max in diameter. The results were so hilarious I feel I absolutely must share them.

The trigger

I recently started Postcrossing. Postcrossing is a worldwide community of people sending each other postcards. I was made aware of its existence after someone on EuroBillTracker linked his/her profile there and I found the idea appealing. It's an activity that's, for now at least, quite cheap, in part thanks to the low postage prices in Germany compared to some other countries, and you get to mail people around the world. Were I in the UK where postage for international postcards seems to be something around 3-4£ (that's almost 5€!!) I don't know if I would've started with it, but here it's just 1.25€ to everywhere (0.95€ for domestic mail).

I got a card from Southern Germany, and it was special in more ways than one: First, it was a Postcrossing meetup where postcrossers would come and write postcards together. Two, all participants "signed" with what looks like personalized stamps: A little pictogram in the middle, and their Postcrossing username around it. What a brilliant idea! I wanted one, too!

So I got to thinking: What if my pictogram would be a finalized golf swing – but the ball is a postcard? And can I use some AI to generate that pictogram? I asked a plethora of models, since my Kagi subscription gives me access to currently 42 of them, including all models big and small from Anthropic, OpenAI and Google. The results are.. not quite what I envisioned. :D

The prompt

Can you generate a very small pictogram with a golfer swinging a club, point of view from behind, but the ball is a mail envelope?

What I was looking for was something akin to this: golfswing final phase Look from behind in the final phase of the swing, and a postcard flying off.

Kagi Quick

Kagi Quick does not generate images, but I tried it nonetheless. As a substitute it generated ASCII art.

  o
 /│\
/ │ \
 ╱ ╲
⌐   ¬

─┼─▶  ✉ ─ ─ ─ 📭
╱ ╲

Or a "slightly more compacted version".

  o
 /|\
 / \
 |
─┼──~~►  ✉
/ \

Or a "minimal one-line style".

⛳ ─~~► ✉

It even advised me on how to read it. How nice!

How to read it:

   · The o / head and /|\ body represent the golfer seen from behind.
   · The ~~► is the club swing arc heading toward the ball.
   · The ✉ is the mail envelope serving as the golf ball.

.. and this is just the beginning!

OpenAI GPT OSS 120B

You can hold a club like that, but swinging it..? I especially like the position of the envelope, like, you're supposed to swing at balls when near the face of others.

Google Gemma 4 31B

We're getting there. Slowly. The ball is on the ground now and we seem to swing the club. With one arm. That has an arm. That seems to "nudge" the ball into the air. Or maybe it's just asking it nicely to fly off. Who knows.

Anthropic Claude 4.5 Haiku

One more tiny step in the right direction. We're now supposed to pose like a gorilla when swinging. This will intimidate the ball and if you then swing with the tip of your fingers it will make you a longhitter.

The model answered outright that it cannot generate images.

DeepSeek V4 Flash

I love the dashed line that's supposed to indicate a swing. But we're still in gorilla mode.

Kagi Research

In the end, I used Kagi's Research AI model which can definitely generate and edit images. After some back and forth the end result was this: pictogram of a golfer in the final swing phase swinging at a postcard-shaped golfball And this one will now be used as my stamp. I think it's a nice combination for a personalized stamp. I also ordered ink of various colors with it.

Small disclaimer

Looking at the exact models one can definitely say that the results are expected: I didn't use any flagship versions of them, and those can definitely also generate images. The result, though, is funny nonetheless. :)