MNiST-IMG-390k

Sumary

Task: Number-To-Image
Dataset: ylecun/mnist
Total training time: ~10 minutes
Inputs: Number (0-9) 
Outputs: 32x32 image
Params: ~391k
Framework: PyTorch, diffusers
Author: Paul Courneya (Harley-ml)

Description

MNiST-IMG-390k is an ~390k parameter model trained to generate an image based on an input number (0-9).

Architecture

Parameter	Value
image_size	32
in_channels	1
out_channels	1
num_classes	10
block_out_channels	[12, 16, 20]
layers_per_block	8
norm_num_groups	4

Tiny diffusion gremlin architecture. Compact enough to run on mortal hardware instead of a datacenter powered by melting glaciers 🫠

Training

Hardware

MNiST-IMG was trained on Google Colaboratory (NVIDA Tesla T4) for ~10 minutes with a batch size of 64 for 10 epochs.

Dataset

ylecun/mnist

Training Results

Loss ended at ~0.39.

Note: I can't provide the raw training logs as I loss it somehwere after training. Sorry!

Generation Examples

At 1000 decoding steps:

At 200 decoding steps:

Inference

Use the script in the repo. inference.py

Related Models

1. SupraMNST-IMG-200k

Citation

@misc{mnist-img-390k,
  title     = {MNIST-IMG-390k: a Tiny Diffusion Model for Generating Handwritten Digits},
  author    = {Paul Courneya; Harley-ml},
  year      = {2026},
  url       = {https://huggingface.co/Harley-ml/MNIST-IMG-390k}
}