vit-face-real-vs-fake

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

vit-face-real-vs-fake

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the desireemcv/face-real-vs-fake dataset. It achieves the following results on the evaluation set:

• Loss: 0.2046
• Accuracy: 0.9467

Model description

More information neeThis model is a fine-tuned Vision Transformer (ViT-Base) for binary image classification of human faces as either REAL (authentic photographs) or FAKE (AI-generated). It was fine-tuned from google/vit-base-patch16-224-in21k on the Computer Vision Face Classification Dataset (v2.0, 2026), which contains facial images sourced from Unsplash (real) and various AI generation methods (fake).

The model achieves 94.67% accuracy on the test set, with balanced performance across both classes (F1=0.94 for FAKE, F1=0.95 for REAL). It was trained for 4 epochs using PyTorch and the Hugging Face Transformers library on a Google Colab T4 GPU.ded

Intended uses & limitations

This model is intended for detecting AI-generated facial images (deepfakes) in binary classification scenarios. It can be used as a baseline for deepfake detection research or as a component in content authenticity verification pipelines.

Limitations:

• Trained on a specific set of AI generation methods; may not generalize to newer deepfake techniques.
• Approximately 14% of original dataset images were unavailable during download, reducing training data.
• Not evaluated against adversarial post-processing (compression, cropping, filters).
• Designed for frontal face images only; performance may degrade on non-frontal or occluded faces.

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 4
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 5.9.0
• Pytorch 2.10.0+cu128
• Datasets 4.8.5
• Tokenizers 0.22.2