Patch-siamese decision layer (reference-based print QC, synthetic prototype)
Trained on johko/print-sheet-qc-pairs. Input: two 96x96 RGB patches (reference patch, camera patch, same frame). Output: single logit; defect if sigmoid(logit) > 0.5.
Validation (unseen sheets, 1080 patches): accuracy 0.8157, AUROC 0.8794.
Synthetic-data prototype - expect a domain gap on real camera images.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support