Struggling to Reproduce Realism in Z-Image Turbo LoRA Training

Wonderful-Reserve728 · reddit · 2026-07-13

This post discusses why it is so difficult to reproduce that initial "photorealistic" effect when doing LoRA training with Z-Image Turbo.

The author explains:

Previously, they got great results using SimpleTuner with a set of configurations that have since been modified and lost;
After trying about 12 different LoRA configurations—including adjustments to rank, learning rate, dropout, and datasets—they still couldn't reproduce it;
Switching to AI Toolkit slightly improved image quality, but it still fell far short of the initial effect.

They also noted a few additional observations:

Some LoRAs require very strong "realistic" prompts during inference to yield good results;
However, this isn't universal; the original version hardly needed a pile of keywords;
They don't want a prompt template where they have to copy-paste 60 words every time, preferring the model to naturally generate results similar to the original.

The post concludes by asking the community for help, suspecting the issue might stem from the dataset, training settings, captions, overfitting, or underfitting.

Original post →

Struggling to Reproduce Realism in Z-Image Turbo LoRA Training

More from Multimodal