---
license: creativeml-openrail-m
tags:
- machine-unlearning
- stable-diffusion
- text-to-image
- image-classification
datasets:
- OPTML-Group/UnlearnCanvas
---

# UnlearnCanvas checkpoints

These are the checkpoints used in the paper [UnlearnCanvas: A Stylized Image Dataset to Benchmark Machine Unlearning for Diffusion Models](https://arxiv.org/abs/2402.11846) (NeurIPS 2024 Datasets and Benchmarks Track). The code and usage instructions are in the [UnlearnCanvas GitHub repository](https://github.com/OPTML-Group/UnlearnCanvas). The dataset is at [OPTML-Group/UnlearnCanvas](https://huggingface.co/datasets/OPTML-Group/UnlearnCanvas).

## Contents

| Path | Description |
|---|---|
| `diffusion/compvis/compvis.ckpt` | Stable Diffusion v1.5 fine-tuned on UnlearnCanvas, in the CompVis (PyTorch Lightning) format. Saved at global step 7000. |
| `diffusion/diffuser/style50/` | The fine-tuned Stable Diffusion model in the diffusers format. It loads with `StableDiffusionPipeline.from_pretrained`. The `unet` folder has an fp32 safetensors file and an fp16 `.bin` file. The `unet_ema` folder holds the EMA weights. |
| `cls_model/style50.pth` | ViT-Large/16 style classifier with 51 output classes. |
| `cls_model/style50_cls.pth` | ViT-Large/16 object classifier with 20 output classes. |
| `style_loss_vgg/` | Pretrained networks for the style loss used in the evaluation: `vgg_normalised.pth`, plus the StyTr2 `transformer`, `decoder` and `embedding` weights at iteration 160000. |

The classifier files are full training checkpoints. Each one is a dictionary with the keys `model_state_dict`, `optimizer_state_dict`, `scheduler_state_dict`, `epoch`, and the train and test accuracy. Load the weights with `torch.load(path)["model_state_dict"]`.

## License

The diffusion checkpoints are fine-tuned from Stable Diffusion v1.5, so they are released under the [CreativeML OpenRAIL-M license](https://huggingface.co/spaces/CompVis/stable-diffusion-license), including its use restrictions.

## Citation

```bibtex
@article{zhang2024unlearncanvas,
  title={UnlearnCanvas: A Stylized Image Dataset to Benchmark Machine Unlearning for Diffusion Models},
  author={Zhang, Yihua and Fan, Chongyu and Zhang, Yimeng and Yao, Yuguang and Jia, Jinghan and Liu, Jiancheng and Zhang, Gaoyuan and Liu, Gaowen and Kompella, Ramana and Liu, Xiaoming and Liu, Sijia},
  journal={arXiv preprint arXiv:2402.11846},
  year={2024}
}
```