whistleroosh's picture
Update README.md
d091fa8 verified
|
Raw History Blame Contribute Delete
2.14 kB
---
library_name: pytorch
tags:
- diffusion
- diffusion-models
- mnist
- cifar10
- edm
- edm2
license: other
---
# Diffusion thesis models
Model weights for the bachelor's thesis:
**Score-Based Diffusion Models: Theoretical Foundations and Optimal Sampling Schedules**
Łukasz Leszko, Mateusz Scharmach
University of Warsaw, Faculty of Mathematics, Informatics and Mechanics, 2026
Code: [github.com/leszkolukasz/diffusers](https://github.com/leszkolukasz/diffusers)
## Structure
| Folder | Thesis experiment | Dataset | Contents |
|---|---|---|---|
| `exp1/` | Experiment 1: Prediction Target | MNIST | `noise`, `score` and `x0` predictors |
| `exp2/` | Experiment 2: Same network, different solvers | CIFAR-10 | EDM medium model |
| `exp3/` | Experiment 3: Same network, different reverse processes | CIFAR-10 | EDM medium model |
| `exp7/` | Pattern-dataset experiments | 2D patterns | Gaussian 8×8, 8×4, 6×6, Spiral5, Beehive |
| `edm2/` | External reference | ImageNet-64 | Official NVIDIA EDM2-S checkpoint |
## File types
For each experiment:
- files without prefix — final EMA/inference weights,
- `raw_*` — raw model weights,
- `trainer_*` — training state (epoch, optimizer and scaler).
## Usage
Download a checkpoint:
```python
from huggingface_hub import hf_hub_download
path = hf_hub_download(
"whistleroosh/diffusion-thesis-models",
"exp2/exp2_x0_norm_x_predictor_edm_medium_cifar10.pth",
)
```
Load it with the thesis code:
```bash
uv run main.py generate \
--model-path "$path" \
--schedule <schedule> \
--sampling-schedule <sampling-schedule> \
--equation <equation> \
--solver <solver> \
--n-steps 32 \
--n-samples 1
```
## EDM2
`edm2/edm2-img64-s-1073741-0.075.pkl` is NVIDIA's official EDM2
ImageNet-64 S checkpoint (`edm2-img64-s-fid`, FID 1.58). It was not trained
by us. See the [EDM2 repository](https://github.com/NVlabs/edm2).
## License
The thesis checkpoints are released under the MIT license.
The EDM2 checkpoint is NVIDIA's official model and is covered by
CC BY-NC-SA 4.0. It is included here for non-commercial research use.