Add model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: pytorch
|
| 3 |
+
tags:
|
| 4 |
+
- diffusion
|
| 5 |
+
- diffusion-models
|
| 6 |
+
- mnist
|
| 7 |
+
- cifar10
|
| 8 |
+
- imagenet64
|
| 9 |
+
- edm
|
| 10 |
+
- edm2
|
| 11 |
+
license: other
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# Diffusion thesis models
|
| 15 |
+
|
| 16 |
+
Checkpoints from experiments on prediction targets, reverse equations,
|
| 17 |
+
samplers, noise schedules, AYS and EDM2.
|
| 18 |
+
|
| 19 |
+
## Contents
|
| 20 |
+
|
| 21 |
+
### `models_final/`
|
| 22 |
+
|
| 23 |
+
Final checkpoints for experiments 1–3:
|
| 24 |
+
|
| 25 |
+
- files without prefix — final EMA weights,
|
| 26 |
+
- `raw_*` — raw model weights,
|
| 27 |
+
- `trainer_*` — training state (epoch, optimizer, scaler).
|
| 28 |
+
|
| 29 |
+
Model files contain a `state_dict` and an embedded `config`.
|
| 30 |
+
|
| 31 |
+
### `edm2/`
|
| 32 |
+
|
| 33 |
+
Official NVIDIA EDM2 ImageNet-64 S checkpoint:
|
| 34 |
+
|
| 35 |
+
```text
|
| 36 |
+
edm2/edm2-img64-s-1073741-0.075.pkl
|
| 37 |
+
```
|
| 38 |
+
|
| 39 |
+
This is the `edm2-img64-s-fid` preset (`FID 1.58`). It was not trained
|
| 40 |
+
here; see the [EDM2 repository](https://github.com/NVlabs/edm2).
|
| 41 |
+
|
| 42 |
+
## Usage
|
| 43 |
+
|
| 44 |
+
Download a file with `huggingface_hub`:
|
| 45 |
+
|
| 46 |
+
```python
|
| 47 |
+
from huggingface_hub import hf_hub_download
|
| 48 |
+
|
| 49 |
+
path = hf_hub_download(
|
| 50 |
+
"whistleroosh/diffusion-thesis-models",
|
| 51 |
+
"models_final/exp2_x0_norm_x_predictor_edm_medium_cifar10.pth",
|
| 52 |
+
)
|
| 53 |
+
```
|
| 54 |
+
|
| 55 |
+
Load it with the code from the thesis project:
|
| 56 |
+
|
| 57 |
+
```bash
|
| 58 |
+
uv run main.py generate \
|
| 59 |
+
--model-path "$path" \
|
| 60 |
+
--schedule edm \
|
| 61 |
+
--sampling-schedule edm \
|
| 62 |
+
--equation probability_flow \
|
| 63 |
+
--solver heun \
|
| 64 |
+
--n-steps 32 \
|
| 65 |
+
--n-samples 1
|
| 66 |
+
```
|
| 67 |
+
|
| 68 |
+
## License
|
| 69 |
+
|
| 70 |
+
The thesis checkpoints are released under the MIT license.
|
| 71 |
+
|
| 72 |
+
The EDM2 checkpoint is NVIDIA's official model and is covered by
|
| 73 |
+
CC BY-NC-SA 4.0. It is included here for non-commercial research use.
|