Darian Tomašević¹, Fadi Boutros², Chenhao Lin³, Naser Damer^2,4, Vitomir Štruc⁵, Peter Peer<sup>1

1</sup> University of Ljubljana, Faculty of Computer and Information Science, Ljubljana, Slovenia
² Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany
³ Xi’an Jiaotong University, School of Cyber Science and Engineering, Xi’an, China
⁴ Department of Computer Science, TU Darmstadt, Germany
⁵ University of Ljubljana, Faculty of Electrical Engineering, Ljubljana, Slovenia

This is the official implementation of the ID-Booth framework, which:

 🔥 generates in-the-wild images of consenting identities captured in a constrained environment
 🔥 uses a triplet identity loss to fine-tune Stable Diffusion for identity-consistent yet diverse image generation
 🔥 can augment small-scale datasets to improve their suitability for training face recognition models

conda create -n id-booth python=3.10
conda activate id-booth
pip install -r requirements.txt

To generate images of identities found in the paper, download their fine-tuned ID-Booth LoRA weights. To create your own fine-tuned model with ID-Booth, download the pretrained ArcFace recognition model, and place the weights into the ArcFace_files directory.

To generate images of a desired identity with Stable Diffusion 2.1, use the diffusers library to load the corresponding LoRA weights, which were trained with the ID-Booth framework. The following example generates in-the-wild images of ID_1:

import torch
from diffusers import StableDiffusionPipeline, DDPMScheduler

base_model = "stabilityai/stable-diffusion-2-1-base"
lora_checkpoint =  "Trained_LoRA_Models/ID-Booth/ID_1/checkpoint-31-6400" # Download or train your own

prompt = "face portrait photo of male sks person, city street background"
negative_prompt = "cartoon, render, illustration, painting, drawing, black and white, bad body proportions, landscape"         

pipe = StableDiffusionPipeline.from_pretrained(base_model, torch_dtype=torch.float16).to("cuda:0")      
pipe.scheduler = DDPMScheduler.from_pretrained(base_model, subfolder="scheduler")
pipe.load_lora_weights(lora_checkpoint)
               
image = pipe(prompt=prompt,
              negative_prompt=negative_prompt,
              num_inference_steps=30,
              guidance_scale=5.0).images[0]

image.save(f"ID_1_{prompt}.png")

Results in the paper can be reproduced with data generated by the inference_ID-Booth.py script.

To perform ID-Booth fine-tuning of Stable Diffusion 2.1 on a new identity, you can follow the train_ID-Booth.py script. The training dataset for a desired identity should include a handful of images along with ID embeddings extracted from these images with a pretrained ArcFace recognition model:

FACE_DATASET
└─── ID_1
│   └─── images
│   |       sample_0.png
│   |       sample_1.png
│   |       ...
│   └─── ArcFace_embeds
│           sample_0.pt
│           sample_1.pt
│           ...
└─── ID_2
└─── ...

The required ID embeddings can be extracted with the extract_ArcFace_embeds.py script. Before running train_ID-Booth.py, specify the path to the source folder with identity images in config_train_SD21.py.

For the evaluation of generated synthetic images, we rely on the following repositories:

• dgm-eval to measure quality, fidelity and diversity,
• CR-FIQA to determine the face image quality,
• 6DRepNet to estimate the pitch, yaw and roll of head poses,
• PyEER to analyse identity consistency and separability.

Notebooks and scripts for reproducing the results in the paper can be found in the Evaluation directory, while fine-tuned LoRA weights of different approaches can be downloaded here.

To also evaluate the utility of the produced data, we also use it to train a deep face recognition model, following the train_FR.py script. The performance of these models is then evaluated on state-of-the-art verification benchmarks with test_FR.py.

If you use the code or results from this repository, please cite the ID-Booth paper:

@article{tomasevic2025IDBooth,
  title={{ID-Booth}: Identity-consistent Face Generation with Diffusion Models},
  author={Toma{\v{s}}evi{\'c}, Darian and Boutros, Fadi and Lin, Chenhao and Damer, Naser and {\v{S}}truc, Vitomir and Peer, Peter},
  journal={arXiv preprint arXiv:2504.07392},
  year={2025}
}

Supported in parts by the Slovenian Research and Innovation Agency (ARIS) through the Research Programmes P2-0250 (B) "Metrology and Biometric Systems" and P2--0214 (A) “Computer Vision”, the ARIS Project J2-50065 "DeepFake DAD" and the ARIS Young Researcher Programme.