Best model for Vietnamese documents (High Accuracy vs Speed) - latin.pth vs latin_g2.pth with download_enabled=False ??

[tiaranrh]

Hi Team,

I am using EasyOCR to scan Vietnamese documents which have complex layouts and stacked diacritics.
My environment requires running EasyOCR with download_enabled=False, so I have to manually download the .pth files and place them in the model directory.

My Questions:

1. For a use case where Accuracy is the first priority (specifically for Vietnamese characters) and speed is secondary, should I manually download and use the Gen 1 latin.pth or the Gen 2 latin_g2.pth?
2. Does latin_g2.pth perform better on dense diacritics compared to the standard latin.pth?
3. If I use latin_g2.pth, do I still use craft_mlt_25k.pth as the detector, or is there a newer detector model I need to download manually?

Thank you!