CIFAR-10 DDPM (18.7M params, from scratch)
A small denoising diffusion probabilistic model (DDPM, Ho et al. 2020) trained on CIFAR-10, written from scratch in ~250 lines of plain PyTorch.
Try it in your browser — the ONNX export runs client-side on WebGPU, generating an image in ~2 seconds on your own GPU.
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
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