# sprite32: a 32x32 text-to-sprite diffusion model trained from scratch, open weights (Apache-2.0) — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/a3c8aa14ed319a7e

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-18 03:05 ingested 2026-11-19 02:37

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xz4ern/sprite32_a_32x32_texttosprite_diffusion_model/)

## retained text



```
I trained a little text-to-sprite model from scratch. 32x32, transparent background, open weights.

[Some handpicked samples](https://preview.redd.it/z4wlfgkgba2i1.png?width=1096&format=png&auto=webp&s=5e61fc10d7a94d194c73db6140665cefb10a4f24)

It’s a 143M parameter UNet trained on 660k sprites, icons and emoji. Training took about 20 hours on my 5090. It works directly on the pixels (no VAE) and the output is RGBA, so the background is actually transparent.

Also, it’s completely vibe coded in 5 or 6 sprints over the last year. I put a lot of work into it, but Claude wrote the code.

There are two versions:

* **1.0** is the full model. It’s a little better and you can turn the guidance up or down, but it uses about 9 GB of VRAM (mostly the 4B text encoder) and takes about 0.9s for a single sprite on a 5090
* **1.0 fast** is the one to start with. It’s about 5x faster (0.16s per sprite on a 5090), fits in under 3 GB, and runs fine on a CPU. It’s a 2.3 GB download instead of 8.7 GB. The quality is close.

Try it here: [sprite32.com](https://sprite32.com). It runs on a spare PC in my house, on CPU until my GPU shows up, so be nice to it.

Weights: [huggingface.co/danemarguglio/sprite32](http://huggingface.co/danemarguglio/sprite32)  
Code: [github.com/danemarguglio/sprite32](http://github.com/danemarguglio/sprite32)
```



document fetched 2026-11-19 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xz4ern/sprite32_a_32x32_texttosprite_diffusion_model
