Ink desk · 2026-11-03
Raw-CT scan-volume model moves into a full Paris 4 training run
A contributor reported completing a revised three-dimensional scan-volume architecture that takes raw CT with global z-score normalization, rather than the altered inputs used by an earlier version. The post says tutorial-set training had been performed and that a full Paris 4 run was underway, with multi-scroll training planned next. A follow-up identifies an attached reference as output from the older augmented-input model. No quantitative comparison or completed full-scroll result accompanies the announcement. Sharper strokes remain an aspiration, not an established improvement, and the attachments have not been independently assessed as evidence of reading accuracy.