domenicor046/reader-v3

Reader V3

Clearer letters on scrolls it has never seen: strokes that Reader v2 breaks apart and other 9 µm models blur into texture or speckle come out whole and separate, with over 10x less false ink than Hecate, and the fastest of them all.

Reader V3 is the upgrade to Reader v2: a drop-in checkpoint for the Vesuvius Challenge team's own ink_9um inference command (koine_machines.inference.infer), for the native 9.36 µm and 8.64 µm scans. Every result is on data the model never trained on.

python -m koine_machines.inference.infer <surface_volume.zarr> reader-v3.pth out.tif --overlap 0.5 --blend-mode hann --direction both

Full results, whole never-seen segments, U08 on PHerc0343 (letters found by my own pipeline, shown for the first time anywhere), the head-to-head and a script that checks every number: https://github.com/DomRusso2/reader-v3

MIT licence. Author: Domenico Russo.