---
license: mit
pipeline_tag: unconditional-image-generation
tags:
- diffusion
- ddpm
- cifar10
- onnx
- webgpu
datasets:
- uoft-cs/cifar10
---

# CIFAR-10 DDPM (18.7M params, from scratch)

A small denoising diffusion probabilistic model ([DDPM, Ho et al. 2020](https://arxiv.org/abs/2006.11239)) trained on CIFAR-10, written from scratch in ~250 lines of plain PyTorch.

**[Try it in your browser](https://huggingface.co/spaces/PeterRabbit/cifar10-ddpm-demo)** — the ONNX export runs client-side on WebGPU, generating an image in ~2 seconds on your own GPU.

![samples](samples.png)

*Uncurated 8×8 grid after 300 epochs.*

## Model details

- U-Net, **18.7M parameters**: three resolution levels (32→16→8), residual blocks with sinusoidal timestep embeddings, self-attention at 16×16 and 8×8, dropout 0.1
- Linear beta schedule, T=1000, trained with the simplified noise-prediction MSE objective
- EMA of weights (decay 0.9995) used for sampling; DDIM sampling with 50 steps
- Unconditional — no class or text control
- Trained 300 epochs (~117k steps, batch 128) in ~4 hours on a single RTX 3090; final loss ≈ 0.029

## Files

| File | Description |
|---|---|
| `checkpoint.pt` | Full PyTorch checkpoint: raw weights, EMA weights, optimizer state (resumable) |
| `unet.onnx` | EMA weights exported to ONNX (fp32), verified to 4e-6 against PyTorch |
| `train_diffusion.py` | Complete training script |
| `sample.py` | Generate image grids from the checkpoint |
| `export_onnx.py` | Reproduce the ONNX export |

## Usage

```bash
pip install torch torchvision
python sample.py --n 64 --seed 42          # sample a grid from checkpoint.pt
python train_diffusion.py --base 128 --epochs 300 --out out_big --sample-every 10 --ema-decay 0.9995   # retrain
```

Source and local web app: [github.com/dannysheesh/cifar10-ddpm](https://github.com/dannysheesh/cifar10-ddpm)