POOPS-DRAW

What is POOPS-DRAW?

It generates terrible images.

Modern LLMs are too good.
Even if you ask them to “draw a terrible picture,” you’ll still get something decent.

POOPS-DRAW prevents that.
No matter how refined your prompt is, the result will be delightfully awful.

But it’s more than just a joke.
It’s also a tool for testing the consistency and intuition of LLMs.
Ironically, the better the model, the more coherent and structured the messy results become — even in this chaotic environment.
Drawing with poops-draw

Drawing with poops-draw

The prompt is: "A frog drinking bubble tea in space."

How it works

You enter a prompt.
The Gemini 2.5 Flash model generates JSON coordinate data.
My humble Linux PC draws the image using Python PIL.
You get the result.

Image quality doesn’t matter here.
It’s terrible. That’s the whole point.

Benchmark

Interestingly, POOPS-DRAW can serve as an LLM benchmarking tool as well.
Powerful models with better reasoning generate more consistent drawings, even under the same messy conditions.

However, high-level reasoning models don’t quite fit the chaotic spirit of this project.
So I focused on models with minimal reasoning capabilities for my benchmarks.

Here’s the key question I wanted to explore:

How well can an AI "draw" using only JSON coordinates?
Spoiler: Stronger models still draw... slightly better garbage.

🎨 View the benchmark of the worst AI art ever

Why did I make POOPS-DRAW?

Reason 1: I needed a terrible image

One day, I needed an absolutely awful drawing.

“Please draw a picture of a dog with really bad drawing skills.” (Actual prompt)

But GPT was too good.
Not just GPT — Gemini, Midjourney, you name it — all models were just too competent.
Ironically, to get a truly bad drawing, I needed advanced prompt engineering.

A terribly drawn dog by GPT

Reason 2: I wanted to test something CPU-based

All the AI news says the same thing:

“Without a GPU, you can’t do anything.”

Fast computation, parallel processing, image generation — it all seems owned by GPUs.

Then I read about Microsoft’s BitNet LLM.
A model built using only -1, 0, and 1.
That article blew my mind.
It proved that heavy resources aren’t necessarily required for building intelligence.

So I thought:

“In an era where fast, high-quality image generation is the norm,
what would happen if I made something slow, janky, and CPU-based?”

It felt like a small, playful rebellion against the AI world.
I didn’t care if it was slow or ugly.
What mattered was proving that a CPU, using just coordinates from an LLM, could complete a drawing.

And that drawing engine became POOPS-DRAW.

Reason 3: Fun and a unique artistic style

Remember the GPT-generated Ghibli-style portraits that went viral this year?
People flooded social media with Ghibli-style profile pics.

AI can create nearly any image now —
but when an image has a recognizable “style,” it feels special.

That got me thinking:

“What if my site’s awful drawing style…
became its own recognizable aesthetic?”

Sure, POOPS-DRAW only allows 2 requests per day due to API token limits.
But even that limitation adds to its slow, weird charm.

If people can laugh at, be inspired by, or even appreciate POOPS-DRAW’s bizarre visual style…
then this silly little project will have been totally worth it.

Well… this might just be the reason I came up with afterward. 😂
But hey, isn’t that reason enough?