---
license: creativeml-openrail-m
base_model: runwayml/stable-diffusion-v1-5
tags:
- stable-diffusion
- diffusion-models
- machine-unlearning
- unlearncanvas
---

# FADE UnlearnCanvas -- SalUn unlearned checkpoints

Stable Diffusion checkpoints after applying **Saliency Unlearning (SalUn)**
([SalUn reference implementation](https://github.com/OPTML-Group/Unlearn-Saliency)) to unlearn a
single artistic style from the UnlearnCanvas full (50-style) checkpoint. One
checkpoint per forgotten style.

## Files

| File | Description |
|---|---|
| `salun_monet.safetensors` | Monet forgotten |
| `salun_picasso.safetensors` | Picasso forgotten |
| `salun_van_gogh.safetensors` | Van Gogh forgotten |

## Evaluation

These checkpoints were evaluated with FADE against the retain-only oracle checkpoints
(see `sungjuncho/fade-unlearncanvas-retain-seed{0,1,2}`) and with UnlearnCanvas's own
accuracy/FID metrics. See the project repo for numbers and evaluation scripts.


## About FADE

FADE (Functional Alignment for Distributional Equivalence) is a metric for evaluating machine
unlearning in diffusion models: it generates images from an unlearned model and a
retain-only oracle, then measures a variational upper bound on the KL divergence
between the two models' output distributions on those images. See the project repo
for the full method and evaluation code: the project's public GitHub repo (link forthcoming)

This checkpoint is a fine-tune of Stable Diffusion v1.5, built on the
[UnlearnCanvas](https://github.com/OPTML-Group/UnlearnCanvas) dataset/benchmark
(Zhang et al., 2024). It inherits Stable Diffusion's CreativeML OpenRAIL-M license --
see the license file for the specific use-based restrictions that apply.

## Citation

```
@article{cho2025referencespecific,
  title={Reference-Specific Unlearning Metrics Can Hide the Truth: A Reality Check},
  author={Cho, Sungjun and Hwang, Dasol and Sala, Frederic and Hwang, Sangheum and Cho, Kyunghyun and Cha, Sungmin},
  journal={arXiv preprint arXiv:2510.12981},
  year={2025}
}
```

## Loading

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

state_dict = load_file("<downloaded_file>.safetensors")
# merge into a CompVis-format LDM model, e.g. via UnlearnCanvas's
# ldm.util.instantiate_from_config, then:
# model.load_state_dict(state_dict, strict=False)
```