PDM-Pure: Effective Purification in One Simple Python Script
PDM-Pure is from our paper: Pixel is a Barrier: Rethinking Adversarial-Based Protection for Diffusion-Based Generation [coming soon]
TODO
- [ ] Patch-based PDM-Pure to fit all resolutions
- [ ] Make it flexible to change to strength
- [ ] Release our paper
Introduction
It is a simple but effective purification framework to purify protective perturbations, all you need to do is to run the single file:
python pdm_pure.py --image [IMAGE_PATH] --save_path [SAVE_PATH] e.g. python pdm_pure.py --image demo/advdm/original.png --save_path demo/advdm/
Installation ⚙️
For installaiton please refer to DeepFloyd IF, which is the strongest opensource pixel space diffusion model (PDM) we used to purify the images.
Results 🔍
Protected by AdvDM [1]
Protected by Mist [2]
Protected by Glaze [3]
Protected by SDS [4]
Reference
[1] Liang, Chumeng, et al. "Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples." International Conference on Machine Learning. PMLR, 2023.
[2] Liang, Chumeng, and Xiaoyu Wu. "Mist: Towards improved adversarial examples for diffusion models." arXiv preprint arXiv:2305.12683 (2023).
[3] Shan, Shawn, et al. "Glaze: Protecting Artists from Style Mimicry by {Text-to-Image} Models." 32nd USENIX Security Symposium (USENIX Security 23). 2023.
[4] Xue, Haotian, et al. "Toward effective protection against diffusion-based mimicry through score distillation." The Twelfth International Conference on Learning Representations. 2023.