Inpaint
Inpainting had the weakest recorded category F1 in this small reference-model check. This metric alone does not establish the cause.
TAFE-ID / 04 / FAILURE LOG
The weak spots
stay visible.
A forensic tool is only useful when its uncertainty is part of the interface. These are the failure modes recorded during the project.
HELD-OUT PROFILE
Some manipulation families remain difficult.
Insert
Insertion also had low reference-model F1. The subset is too small to isolate the causes or establish a reliable category ranking.
GroupNorm pilot
The GroupNorm baseline reached only 0.0364 F1 at update 1000 on its held-out subset. Strong fixed-crop scores did not translate into useful document localization.
LIMITS
What the demo does not claim.
The public page does not claim a newly trained state-of-the-art model, a complete paper benchmark reproduction, or production readiness. It exposes a verified reference inference path and the evidence needed to reproduce the decision.