Krea 2 LoRA training: Raw vs Turbo, and which trainer actually works?

Sweaty_Painting1978 · reddit · 2026-09-29

A Redditor wants to train LoRAs on Krea 2, which they call the best open-source image model right now, but finds contradictory advice: some say train on Raw and generate on Turbo, others add Turbo adapter training. A default-settings AI Toolkit run at 1024 resolution underperformed, and the tool only supports one GPU per run, so they're weighing AI Toolkit vs Musubi Tuner vs Fizgig against their usual kohya sd-scripts multi-GPU setup. They also ask how natural-language captions for Krea 2 should be written—length, style, trigger word handling.

Original post →

More from Multimodal

Multimodal channel →