FADM - Full Body Anonymization Using Diffusion Models

Introduction

This is the official implementation for the paper "Context-Aware Full Body Anonymization using Text-to-Image Diffusion Models" presented at the ACVR 2024 workshop in conjunction with ECCV 2024.

Getting started

• Clone this repository
• create a new conda environment
• pip install -r requirements.txt
• run setup.py

Example Usage

• All arguments are availbale using python anonymize.py --help
• Single file python anonymize.py --file example.jpg
• Folder python anonymize.py --input path/to/folder/ --output path/to/folder/
• Add --out_sbs to output a side by side image (original left, anonymized right)
• Choose the model version by using --version v : $v \in {\text{21, 20ip, rv6ip}}$ standing for SD 2.1, SD 2.0 inpainting and Realistic Vision 6 Inpainting
• --ds 1.0 shows interesting results
• --no_pose seems to work
• Use --highvram to enable model caching on the GPU when enough VRAM is available

During the first run, the program will download the needed models and dependencies!

Multi GPU support

Only one GPU can be used at the moment, but running multiple instances to anonymize different folders in parallel can be done by adding --cuda_device N to run on GPU N

Acknowledgement

This project emerged during the research projects ANYMOS - Competence Cluster Anonymization for networked mobility systems (grant number 16KISA085K) and just better DATA (jbDATA) supported by the German Federal Ministry for Economic Affairs and Climate Action of Germany (BMWK) and was founded by the German Federal Ministry of Education and Research (BMBF) as part of NextGenerationEU of the European Union.