Summary

This project provides the official PyTorch implementation for Anti-DreamBooth, a defense system against malicious personalized text-to-image synthesis using models like DreamBooth. It targets researchers and users concerned about privacy, offering a method to protect individuals' likenesses from being synthesized into fake or disturbing content, thereby safeguarding personal privacy against AI manipulation.

How It Works

Anti-DreamBooth introduces subtle, optimized noise perturbations to reference images before DreamBooth training. This degrades the quality of synthesized target images, disrupting personalization without significantly altering the original image's utility. The system evaluates various perturbation optimization algorithms for robustness against different text-to-image models and conditions.

Quick Start & Requirements

• Installation: Clone repo, create/activate Conda env (conda create -n anti-dreambooth python=3.9, conda activate anti-dreambooth), install dependencies (pip install -r requirements.txt).
• Prerequisites: Python 3.9, PyTorch, Hugging Face diffusers, ShivamShrirao's diffusers fork. LoRA requires specific versions: diffusers==0.23.1, accelerate==0.33.0, transformers==4.48.3. Pretrained Stable Diffusion checkpoints (v1.5, v1.4, v2.1) needed in ./stable-diffusion/.
• Hardware: Tested on a single NVIDIA 40GB A100 GPU.
• Datasets: VGGFace2, CelebA-HQ used for experiments.
• Links: Official implementation of ICCV 2023 paper "Anti-DreamBooth: Protecting users from personalized text-to-image synthesis".

Highlighted Details

• Official implementation of the ICCV 2023 paper on defending personalized text-to-image synthesis.
• Uses subtle noise perturbations to disrupt DreamBooth training on user images.
• Effective against model/prompt mismatches.
• Supports ASPL, E-ASPL, FSMG, T-FSMG, and E-FSMG defense algorithms.

Maintenance & Community

Maintained by VinAI Research. Contact imthanhlv@gmail.com or use GitHub issues for support. Recent updates include evaluations code and dataset links. No community channels listed.

Licensing & Compatibility

The repository README does not explicitly state the software license. This requires clarification for adoption, particularly for commercial use or integration into closed-source projects.

Limitations & Caveats

Focuses on defending against DreamBooth; effectiveness against other synthesis methods is not detailed. Experimental setup requires high-end GPU hardware (NVIDIA 40GB A100). Absence of a specified license is a significant adoption blocker.