I had a peculiar problem trying to have AI generate a tileset for use in developing 2D games as a placeholder and I think it taught me a bit of a lesson. I think we need to think of how these models handle things beyond the simple benchmarks or the idea that they all do pretty much the same thing.
The Prompt:
"Create a top-down, 32x32 pixel art tileset for a 2D RPG town scene similar to games like a link to the past and final fantasy. Include grass, dirt, pavement, sidewalk, cars and various objects and water tiles. Create a clear, orderly sprite sheet with sharp, 1-pixel outlines and a consistent 16-color palette. Include some city assets as well as a building that says 'About me', a building that says 'Projects' and a building that says 'Get in touch', make the buildings look appropriate to their name."
The Gemini Attempt
I tried gemini (Using nano banana I assume, free account on web app) first and it took a while to get it to make proper perspective that didn't look like it was pulled directly off a major copyrighted work (In this case stardew valley).
Eventually I also found that Gemini seems to be bound somehow to make images across the whole canvas only, unable to specify that we want for example 32x32 pixel tiles in a 512x512 canvas, for example.
The ChatGPT Attempt
I then tried ChatGPT (in free mode, webapp) who had a similar issue, but more readily understood the needed perspective and tiles.
ChatGPT then let me know that it had the same issue when generating individual parts within a canvas, and could not ensure that my canvas would be the size I requested.
The Grok Attempt (for what it's worth)
Grok (in free mode on webapp) generated multiple blank images and then informed me that I had reached my quota. Nice.
The Z.ai Anomaly
Then I tried z.ai (free, using GLM 5 Turbo) and it immediately set out to do things differently by creating a site where each tile was being generated pixel by pixel.
And while I'll admit that the images produced were much more rudimentary they were the only tiles which were indeed 32x32 pixels.
DISCLAIMER: I got this HTML code from the 'thinking' section as I think the free token limit was hit or something because there was no output outside of thinking. Furthermore the site did not actually work, and the color palette also did nothing. You be the judge as to whether this should disqualify z.ai or whether that even addresses the difference in approach here.
In conclusion - I saw some 'smart' models handle the problem by attempting to brute force methods which were ineffective be beaten by a more elegant thought process.