Technical architecture

Pixel-only forensic branches around a controlled semantic backbone.

M4 and M5 isolate two different branch additions. The full three-branch system remains a future experiment and has not been evaluated.

384×384 RGB
portrait + context crop

SemanticConvNeXt-Tinyhigh-level representation

Texture32×32 → 16→32→6464-D representation

EdgeLuminance → Sobel X/Y32-D representation

Tested as two-branch systems
M4: Semantic + Texture · M5: Semantic + Edge

Semantic

ImageNet-pretrained ConvNeXt-Tiny at 384×384. Stem and stages 0–2 are frozen; stage 3 and final normalization are trainable.

Texture · M4

Adaptive 32×32 pooling, Conv-BN-ReLU at 16→32→64 channels, global pooling, and a 64-D representation.

Edge · M5

Luminance grayscale, fixed canonical Sobel X/Y, stable gradient magnitude, learned 1→16→32 convolutions, and a 32-D representation.

Common-space fusion

Each tested forensic feature and the semantic embedding are independently projected to 256-D. Two zero-initialized logits are softmax-normalized, then the weighted features are added and classified through 256→512→256→128→3.

Best M4 weights: semantic 0.503779, texture 0.496221. Best M5 weights: semantic 0.502496, edge 0.497504. Learned fusion weights are not causal feature-importance estimates.

Pixels only

No detector identity, confidence, padding flag, acquisition device, presentation condition, file path, or source metadata enters a classifier.