# I trained a tiny image model from scratch that draws newsprint faces with custom text — 2–4 steps, CPU-only — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/c599346b673353bb

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-02 13:53 ingested 2026-11-03 02:40

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

## retained text



```
hiii! I made grinpress v3, a tiny 128x128 rectified-flow model that draws newsprint-style faces with words beside them :3



It's trained from scratch, with character-level text conditioning. You can choose seven expressions, colors, paper styles, layouts, and text up to 14 characters, including mixed case and punctuation.



For v3, I distilled the 50-step teacher into 2- and 4-step students and exported them to ONNX with dynamic INT8 quantization.



CPU inference measured approximately 0.27 seconds at 2 steps / 0.54 seconds at 4 steps on my benchmark's 8-vCPU Intel Xeon VM, using 8 threads after warm-up. Other CPUs can be slower.



It's a small, specific domain: newsprint faces and lettering at 128x128. The weights are MIT licensed, and there's a live CPU demo. Long phrases can get cramped, and it occasionally mangles a character or word spacing.



Try it: [https://huggingface.co/spaces/skyuu72/grinpress-v3-demo](https://huggingface.co/spaces/skyuu72/grinpress-v3-demo)



Weights, benchmarks, and limitations: [https://huggingface.co/skyuu72/grinpress-v3](https://huggingface.co/skyuu72/grinpress-v3)



I'd love to see which words break it, or hear how it runs on your CPU. tiny newspaper creature feedback welcome uwu :3
```



document fetched 2026-11-03 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xlxrqz/i_trained_a_tiny_image_model_from_scratch_that
