Training LoRA for Krea 2 Turbo w/Training Adapter with AI-Toolkit
I wanted to share my experience because I couldn't find many reports about training Krea 2 Turbo LoRAs on a 12GB GPU. This was my first successful Krea 2 Turbo LoRA after some trial and error.
GPU: RTX 5070 12GB VRAM
RAM: 64 GB
OS: Windows 11
dataset:
36 1024x1024 images
Realistic female character
Rank:32
Resolution: 1024
LoRA for Kerea 2 Turbo w/Training Adapter
cache_latents_to_disk: true
cache_text_embeddings: true
Initially training appeared to hang. It was actually creating latent cache and text embeddings. After waiting, training started normally.
Changing
lokr_full_rank: true
to
lokr_full_rank: false
reduced step time from 25.4 sec to 22.3 sec.
Training time: 18h45m
LoRA size: 223MB
Training completed successfully. Character identity is generally consistent across generations. Body consistency is very good. Face consistency is acceptable. Some generations still deviate from the target face, but complete identity drift is rare.
Initially I thought 12GB VRAM would be insufficient. However,
cache_latents_to_disk: true
cache_text_embeddings: true
lokr_full_rank: false
made training practical.
Hopefully this helps someone who is wondering whether 12GB VRAM is enough.
Thanks!
Thank you for sharing! curious as to why you chose to train on Turbo + Turbo LoRA instead of directly training on Raw ?
Thank you for your comment. Because, I didn't think 12GB of VRAM was sufficient for LoRA training based on raw. However, a recent Reddit post shows someone doing exactly it, so I will try it myself.