Synthetic Sight

Detecting AI-generated faces for digital literacy

AI4ALL Ignite · Summer 2026 · Team 9C

Julissa Lema · Kristine Johnson · Ricky Dixon · Shloka Kandukuri

Quick start

git clone https://github.com/kristinealli/SyntheticSight.git
cd SyntheticSight
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

python -m src.inference --image path/to/image.jpg

Documentation

• Data card — dataset source, preprocessing, splits, and limits
• Evaluation — metrics, threshold selection, and errors
• Bias and ethics — audit scope and responsible use
• Project history — consolidation and checkpoint audit

Synthetic Sight is a reproducible computer-vision research prototype that asks:

Can a CNN classifier distinguish real human faces from StyleGAN-generated synthetic faces well enough to help flag possible misinformation?

The final system uses ImageNet-pretrained ResNet-50 transfer learning to classify still face images as Real (0) or Synthetic/Fake (1).

Synthetic Sight is a review-support signal, not an authentication service or a universal deepfake detector.

Held-Out Benchmark Result

Metric	Final result
Locked test images	20,000 (10,000 real / 10,000 synthetic)
Accuracy	99.69%
Fake F1	99.69%
ROC-AUC	99.995%
Decision threshold	0.51
False positives	39 real images flagged synthetic
False negatives	23 synthetic images labeled real

For evaluation details, including class-specific metrics and the confusion matrix, see docs/evaluation.md.

These results show extremely strong separation within this FFHQ-versus-StyleGAN benchmark. They do not establish equivalent performance for newer generators, face swaps, video, screenshots, recompressed media, edited real images, or uncontrolled real-world images.

The test split was held out while the model, checkpoint, and threshold were developed. These results show extremely strong separation within this benchmark; they do not establish the same performance on newer generators, face swaps, video, screenshots, heavy recompression, or uncontrolled real-world images.

Repository history

This repository is consolidated around the final ResNet-50 pipeline used for the presentation. See docs/project-history.md for the full audit trail and earlier experiments.

Method

flowchart LR
    A[RGB face image] --> B[Resize 224 × 224]
    B --> C[ImageNet normalization]
    C --> D[ResNet-50 backbone]
    D --> E[Global average pooling]
    E --> F[Linear 2048 → 256]
    F --> G[BatchNorm + ReLU + Dropout 0.30]
    G --> H[Linear 256 → 1]
    H --> I[Sigmoid synthetic score]
    I --> J{score ≥ 0.51?}
    J -->|Yes| K[Synthetic / Fake]
    J -->|No| L[Real]

Loading

Final training configuration

• Dataset: Kaggle 140k Real and Fake Faces — FFHQ real faces + StyleGAN synthetic faces.
• Splits used: 100,000 train, 20,000 validation, 20,000 locked test images.
• Input: RGB, 224 × 224, ImageNet normalization.
• Training augmentation only: horizontal flip (p=0.5) and light color jitter.
• Stage 1: freeze the pretrained backbone and train the new classifier head for 3 epochs.
• Stage 2: unfreeze ResNet-50 layer4 and fine-tune it with the head.
• Loss: BCEWithLogitsLoss.
• Optimizer: Adam, weight decay 1e-4.
• Learning rates: head 5e-4; layer4 1e-5 when fine-tuning starts.
• Seed: 13; batch size 32; maximum 15 epochs.
• Threshold selection: validation sweep from 0.05 to 0.95 in 0.01 increments, maximizing Fake F1.
• Selected model: epoch 13 at threshold 0.51.

The final notebook's implementation monitors validation Fake F1 for checkpointing. Epoch 13 also has the lowest recorded validation loss, which is why the final presentation's “lowest validation loss” description identifies the same checkpoint.

Repository layout

SyntheticSight/
├── README.md
├── assets/                  # Final-run plots used by documentation
├── deployment/
│   ├── streamlit_app.py     # Interactive research prototype
│   ├── api.py               # FastAPI inference endpoint
│   └── Dockerfile
├── docs/
├── models/
├── notebooks/
├── scripts/
├── src/synthetic_sight/     # Shared model and inference code
└── tests/                   # Preprocessing/model contract tests

Setup

1. Create an environment

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
python -m pip install --upgrade pip
pip install -e ".[app,api]"

For notebook work:

pip install -e ".[research,dev]"

2. Verify the included final checkpoint

The verified epoch-13 final checkpoint is included at:

models/best_resnet50.pth

Verify its architecture and metadata contract before deployment:

python scripts/verify_checkpoint.py models/best_resnet50.pth

Expected SHA-256:

d9a7fd6a692c942b550f9848500dc3ffb10d5809cb0d0091990648bf369ad21c

The validator checks the final architecture contract, epoch, label mapping, and threshold. A legacy artifact fails rather than silently producing mismatched predictions.

3. Run the Streamlit prototype

streamlit run deployment/streamlit_app.py

4. Run the FastAPI service

uvicorn deployment.api:app --reload

Interactive API documentation is available at /docs while the service is running.

Evaluation

The official test set contains 10,000 real and 10,000 synthetic images and was evaluated only after development decisions were complete.

Ground truth	Predicted Real	Predicted Synthetic
Real (10,000)	9,961 TN	39 FP
Synthetic (10,000)	23 FN	9,977 TP

See docs/evaluation.md for metric definitions, threshold interpretation, and error analysis.

Bias, representation, and responsible use

Balanced Real/Fake classes do not establish demographic fairness. The project therefore added a preliminary apparent-lightness audit using OpenCV face detection, forehead/cheek sampling, and CIE Lab L*. This is a measurable image characteristic—not race, ethnicity, gender, ancestry, or inherent skin color—and it is sensitive to lighting, exposure, editing, makeup, and face-detection errors.

The detector should flag content for review, not authenticate it. False positives can cast doubt on legitimate images; false negatives can create false reassurance. See docs/bias-and-ethics.md.

Reproducibility notes

• File paths were checked for zero overlap across training, validation, and test partitions.
• The test split stayed locked until checkpoint and threshold decisions were complete.
• The final training provenance traces to the executed source notebook training_exploration_reinteration_kristine.ipynb; the polished repository includes its consolidated counterpart as notebooks/01_resnet50_final_training.ipynb.
• The notebook records the sample configuration, labels, seed, transforms, training schedule, checkpoint metadata, and evaluation logic.
• The final repository deliberately rejects legacy checkpoints whose architecture metadata does not match the final run.
• A stronger future audit should also check image near-duplicates and source-level correlations.

Planned Next Steps

1. Test the existing model on unseen generator families before retraining to measure generator drift.
2. Evaluate compression, resizing, screenshots, filters, and other real-world transformations.
3. Join apparent-lightness audit rows to prediction/error records and compare error rates with sample sizes and uncertainty.
4. Evaluate probability calibration rather than assuming model scores are literal real-world probabilities.
5. Extend beyond still images with video-specific, spatiotemporal benchmarks and architectures.

Documentation

• Architecture
• Dataset and data protocol
• Evaluation and metrics
• Bias and ethics
• Deployment
• Project evolution
• Repository review and consolidation decisions
• References

References

Core references include He et al. (2016) for ResNet, Deng et al. (2009) for ImageNet, Karras et al. (2019) for StyleGAN, Mo (2020) for the benchmark dataset, and Nightingale & Farid (2022) for the human-perception motivation. Full citations are in docs/references.md.

---

Use statement: Synthetic Sight is an educational research prototype. Its output is a model score under a specific benchmark distribution; it should not be treated as legal proof, identity verification, provenance certification, or an automated high-impact decision.