# I created the First diffusion AI model for the Game Boy. — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/30494487c6a9cd07

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-30 08:35 ingested 2026-10-31 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xj9w8d/i_created_the_first_diffusion_ai_model_for_the/)

## retained text



```
My Game Boy Color takes 85 seconds, and it has to fake every single multiplication with a lookup table because its CPU can’t multiply. But it WORKS!

This is a real neural network running 16 DDIM steps on the handheld’s own 8 MHz CPU, with no PC attached and no streaming. As soon as I press A, it starts from pure noise and a doodle comes out. I was trained on 880,000 doodles, and only the 1.08M learned weights are in the actual GB ROM.

I did not find anyone who got a diffusion model running on a Game Boy before, so I’m calling it the first. If you know of an earlier one, please show me, I’d love to see it!

The numbers:

* MLP denoiser, 784 → 512 → 512 → 784, 1.08M parameters
* Trained on Google’s Quick, Draw!: 8 classes, 110k drawings per class, 28x28
* Int8 weights, 4-bit activations, 16-bit sums
* Classifier-free guidance distilled into the weights, so one forward pass per step
* About 8.8 million multiply-adds per image
* 2 MB ROM, almost all of it weights

The fun part was getting it to fit. A 24 KB product table in WRAM turns every multiply into two lookups, which comes down to 18 machine cycles per multiply-add. Layer 1 overflowed 16 bits, so its weights lost a bit of precision. The training does the ROM’s exact integer math, and an emulator checks the ROM against the Python version after every one of the 16 steps. They match bit for bit, for all 8 categories.

My 17 year old flash cart was a real pain in the ass and often just 'died' when trying to erase it. Old hardware keeps you humble :P.

Is it any good? Ehhh, it’s a Game Boy. Apples sometimes stay open and fish sometimes fall apart, but it’s drawing them from noise on a handheld from 1998, so I’m taking it.

The music is NOT AI but created by me, real human yo.

If you wanna try or contribute: try the ROM or read the code: [https://mowelan.github.io/gb-diffusion/](https://mowelan.github.io/gb-diffusion/)
```



document fetched 2026-10-31 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xj9w8d/i_created_the_first_diffusion_ai_model_for_the
