AI Shield Final Detector
Upload one image. The app runs the frozen forensic branches, the semantic ViT attention branch,
and the final GLM fusion model that combines all evidence into one fake probability.
EfficientNet forensic branch
VAE anomaly branch
Face-gated OpenForensics
Custom semantic attention
GLM late fusion
0 1
Final scores
This is the part to look at for the actual model decisions. The Phase 1 score uses only the forensic layer. The final Phase 3 score uses both the forensic layer and the semantic attention layer.
Final model scores
Branch evidence
The two actual decision scores are shown in the Final Scores tab and in the cards above. These rows are not all the same type of probability. EfficientNet and semantic ViT are fake-probability scores. The VAE row is an anomaly percentile. The face gate is only an applicability check. OpenForensics is neutral at 50% when no valid face is present.
Evidence values
Features sent to final GLM
Phase 3 comparison table