README.md

---
license: mit
tags:
  - diffusion
  - ddpm
  - ddim
  - from-scratch
  - pytorch
  - unconditional-image-generation
pipeline_tag: unconditional-image-generation
---

# diffusion-from-scratch — trained checkpoints

Weights for [github.com/adimunot21/diffusion-from-scratch](https://github.com/adimunot21/diffusion-from-scratch),
a DDPM/DDIM implementation written from scratch in PyTorch — forward process, noise
schedules, U-Net denoiser, samplers and classifier-free guidance, no diffusers dependency.

Three models. Each folder has `config.json` plus two weight files:

- **`ema.safetensors`** — exponential moving average of the weights (`ema_decay=0.9999`). **Use these for sampling.**
- **`model.safetensors`** — the raw final training weights, kept for completeness.

| Folder | Model | Params | Schedule | Epochs | Final train loss | FID |
|---|---|---|---|---|---|---|
| `mnist/` | MNIST, unconditional | 9.53 M | linear, T=1000 | 50 | 0.0210 | 34.1 |
| `cifar10_uncond/` | CIFAR-10, unconditional | 46.03 M | cosine, T=1000 | 100 | 0.0547 | 71.2 |
| `cifar10_cond/` | CIFAR-10, class-conditional | 46.03 M | cosine, T=1000 | 100 | 0.0557 | 65.3 |

The CIFAR-10 U-Nets use channels `[128, 256, 256, 512]`, 2 residual blocks per level and
4-head self-attention at the two middle resolutions. The conditional model was trained
with classifier-free guidance (`uncond_prob=0.1`, 10 classes).

## Loading

These are plain `state_dict`s for the `UNet` class in the source repo — not a
`diffusers` pipeline. Build the model from the repo, then load:

```python
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

path = hf_hub_download("adimunot/diffusion-from-scratch",
                       "cifar10_uncond/ema.safetensors")
model.load_state_dict(load_file(path))   # model = UNet(**config) from the repo
model.eval()
```

`config.json` carries the exact architecture and training hyperparameters each
checkpoint was produced with, so it can be fed straight into the constructor.

## Notes

- FID was computed during the project's evaluation phase; treat the numbers as
  self-reported and comparable only within this repo.
- These are learning-project models trained on a single GPU, not competitive baselines.
  CIFAR-10 FID in the 65–71 range reflects the small model and 100-epoch budget.