Training Krea 2 LoRAs: Can We Reuse Old Tag-Based Datasets?

poliranter · reddit · 2026-08-05

A creator with a massive backlog of SD 1.5 and Pony era assets asked if old datasets trained with Tags could be directly used for Krea 2 models. Since newer models generally prefer natural language descriptions, the poster wants to know the performance difference of reusing old tags to avoid re-captioning massive image libraries.

Original post →

More from Multimodal

Multimodal channel →