buildborderless/CommunityForensics-DeepfakeDet-ViT

NOTICE: 7/14/2026

• We have received multiple reports in the past few months regarding the misuse of this model and its license. For the time being until we can figure out a better solution, the model will temporarily be set to gated mode.
• We encourage all to review the MIT License: the Community Forensics detection model can be forked, distributed, and utilized for use in for-profit operations, but attribution must be given.
• Attribution details can be found below under "Citation" for the model. For the transformers implementation, a simple awknowledgement to Borderless would be appreciated.

Trained on 2.7M samples across 4,803 generators (see Training Data)

Model presented in Community Forensics: Using Thousands of Generators to Train Fake Image Detectors.

Uploaded for community validation as part of OpenSight - An upcoming open-source framework for adaptive deepfake detection.

Project OpenSight HF Spaces coming soon with an eval playground and eventually a leaderboard. Preview:

Model Details

Model Description

Vision Transformer (ViT) model trained on the largest dataset to-date for detecting AI-generated images in forensic applications.

• Developed by: Jeongsoo Park and Andrew Owens, University of Michigan
• Model type: Vision Transformer (ViT-Small)
• License: MIT (compatible with CreativeML OpenRAIL-M referenced in [2411.04125v1.pdf])
• Finetuned from: timm/vit_small_patch16_384.augreg_in21k_ft_in1k
• Adapted for HF inference compatibility by Borderless.

HF Space will be open sourced shortly showcasing various ways to run ultra-fast inference. Make sure to follow us for updates, as we will be releasing a slew of projects in the coming weeks.

Links

• Repository: JeongsooP/Community-Forensics
• Paper: arXiv:2411.04125
• Project Page: https://jespark.net/projects/2024/community_forensics

Training Details

Training Data

• 2.7mil images from 15+ generators, 4600+ models
• Over 1.15TB worth of images

Training Hyperparameters

• Framework: PyTorch 2.0
• Precision: bf16 mixed
• Optimizer: AdamW (lr=5e-5)
• Epochs: 10
• Batch Size: 32

Evaluation

Unverified Testing Results

• Only unverified because we currently lack resources to evaluate a dataset over 1.4T large.

Metric	Value
Accuracy	97.2%
F1 Score	0.968
AUC-ROC	0.992
FP Rate	2.1%

Re-sampled and refined dataset

• Coming soon™

Citation

BibTeX:

@misc{park2024communityforensics,
    title={Community Forensics: Using Thousands of Generators to Train Fake Image Detectors}, 
    author={Jeongsoo Park and Andrew Owens},
    year={2024},
    eprint={2411.04125},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2411.04125}, 
}