PolLabarbarie/diffusion-signflip-anon-face

FFHQ ControlNet for Diffusion Sign-Flip Anonymization

This repository contains the code for the paper "Secure and reversible face anonymization based on a diffusion model with face mask guidance" (to be published) by Pol Labarbarie, Vincent Itier and William Puech. See the github repository for the full code and instructions: (https://github.com/PLabarbarie/diffusion-signflip-anon)

This hugging face repository contains a custom ControlNet checkpoint trained for the diffusion sign-flip anonymization pipeline.

The model is used to condition an FFHQ latent diffusion model during face anonymization and reconstruction. This variant uses segmentation-mask conditioning.

The included ffhq-diffusers/ directory contains the FFHQ diffusion weights from the original paper (https://arxiv.org/abs/2112.10752), converted to match the Hugging Face Diffusers library format.

The repository also includes the precomputed sign-flip keys used by the paper experiments. These keys are provided for reproducibility.

Loading

import os
import torch

from anonymization.diffusion import DiffusionModel
from anonymization.controlnet import ControlNet
from config.config_main import config

device = "cuda"

diffusion_model = DiffusionModel(
    name="ffhq",
    torch_device=device,
    models_root=config.models_root,
    weights_root=config.weights_root,
)

controlnet = ControlNet(
    model_config=diffusion_model.unet.config,
    model_ckpt="ffhq-diffusers",
    hint_channels=3,
    down_sample_factor=4,
    device=device,
)

state_dict = torch.load("controlnet_epoch_15.pth", map_location=device)
controlnet.load_state_dict(state_dict)
controlnet.eval()

Files

controlnet_epoch_15.pth
controlnet_config.json
ffhq-diffusers/
keys/
|-- keys_CelebA_HQ.pt
|-- sub_keys_diversity_0.pt
|-- sub_keys_diversity_1.pt
|-- ...
`-- sub_keys_diversity_9.pt
README.md

Citation

If you use this checkpoint, please cite the associated paper once available.

@article{diffusion_signflip_anonymization,
  title   = {Secure and reversible face anonymization based on a diffusion model with face mask guidance},
  author  = {Pol Labarbarie and Vincent Itier and William Puech},
  journal = {},
  year    = {2026}
}