Exploring LoRA Training Settings for Ideogram 4: 1700 Steps with 24 Images

Zealousideal-Car4724 · reddit · 2026-07-30

A user shared practical experience on training a LoRA model using Ideogram 4. The model is configured with Rank 32 and Alpha 16, utilizing a dataset of 24 images.

The author found that training for 1,300 steps resulted in an underfitted model with many generation errors, but increasing to 1,700 steps showed noticeable improvements. They are now seeking community advice on whether pushing the training to 2,000 or 2,500 steps would yield better results.

Original post →

More from Multimodal

Multimodal channel →