THEIA on a web application
Confusion matrix, error analysis
Recall, precision, F1 curves

THEIA 👁️ by QTPies

(Note: In Greek mythology, Theia is the goddess of sight and shining light: the perfect namesake for a model designed to see through synthetic deception).

Motivation 💡

In the age of viral media, AI-generated images are fueling real harm: from scam product listings to synthetic photos used to impersonate real people during breaking events.

However, we discovered a massive vulnerability in current moderation systems: real-world wear and tear. During our testing, we found that AI-generated images initially flagged by standard detectors completely bypassed detection after simply going through TikTok’s native compression and re-upload process. Standard detectors rely on fragile metadata (like C2PA) or invisible watermarks that vanish after a screenshot or WhatsApp forward.

We built THEIA to be a "Zero-Trust" pixel-level detector. We designed it to hold up after real-world transformation, because catching AI-generated images only matters if the model still works after the image has been through what every file online actually goes through.

Key Features ✨

Media Upload Gateway: Seamlessly upload suspected images or videos.
Live Dashboard Scoring: Intuitive visual breakdowns of the media's authenticity.
Real-Time Probability Engine: Instant, granular confidence scoring p(AI-generated) ∈ [0,1] based on semantic pixel structure, not metadata.

Our Unique Value Proposition 🛡️

Instead of chasing lab-only benchmark scores, we optimized for genuine robustness.

Zero-Trust Metadata: We ignore fragile metadata entirely, focusing only on the robust, structural anomalies left by generative AI.
Scientific Rigor over Hype: During development, we caught two real problems that other teams would have shipped. First, we identified a hidden resolution shortcut that was incorrectly flagging real, low-resolution photos as fake 87% of the time. Second, we actively falsified and retracted a wrong assumption about our own performance gains.
True Generalization: We deliberately shipped the model variant that generalized best to unseen data, rather than the one that simply overfit to the organizer's designated benchmark.

Accomplishments We're Proud Of 🏆

By leveraging a frozen backbone and aggressive stacked-transformation training (compressing, blurring, and injecting noise), THEIA achieves state-of-the-art robustness:

Baseline Accuracy: Clean ROC-AUC of 0.9991.
Transformation Robustness: On our worst single-transform cell (heavy Gaussian noise $\sigma=0.10$), the model held strong with a 0.9951 AUC (a max degradation drop of only 0.0039).
External Generalization: Achieved 0.9917 AUC on completely out-of-distribution, independent test sets.

How We Built It 🛠️

🛢️ Datasets & Leakage Prevention

Training (Reals/Fakes): SID_Set (24 shards), WildFake (LAION reals, Midjourney v5, 6 GAN families), and COCO test2017 reals.
Validation: 4,998 COCO val2017 reals + 8,843 DALL-E 3 fakes (organizer-designated).
External Generalization Sets: Community Forensics-Eval (CVPR 2025, 10+ generators) and RRDataset.
Security: Data leakage was prevented via strict pixel-content hashing, enforced automatically before every single training run.

💻 Tech Stack

Models: google/siglip2-giant-opt-patch16-384 (1.164B base, frozen backbone).
Core ML: PyTorch, scikit-learn, SciPy, NumPy, pandas, torchvision, Hugging Face (Transformers, Accelerate, safetensors).
Image Processing: Pillow, Kornia.
Visualization: Matplotlib, seaborn.

What We Learned & What's Next 🚀

Taking inspiration from recent literature, we realized that data quality and curation matter far more than sheer architectural complexity.

Looking forward, our roadmap includes:

Edge Deployment: Because THEIA is strictly restricted to a <2B parameter model (1.164B), it maintains a highly lightweight memory footprint. Our next step is optimizing the model for on-device mobile deployment (iOS/Android).
Latency & Scalability: Transitioning the Python backend into scalable, production-ready microservices to handle high-throughput, real-time video frame analysis with sub-second latency.
Video Temporal Analysis: Expanding beyond frame-by-frame detection to analyze temporal flickering and inconsistencies across full video files.