Version: 1.0
Trigger Word: midriff
Model Type: LoRA (Z-Image De-Turbo)

MidriffZit is a focused LoRA designed to improve how Z-Image models generate midriff looks where the waistline is visible.
The LoRA is trained on a curated Korean girls dataset.

Training Details

• Dataset theme: Korean fashion/portrait references

• Caption: midriff

• Resolution: 512

• Steps: 3000

• Batch size: 1

• Noise scheduler: FlowMatch

• Optimizer: AdamW 8-bit

• UNet trained, text encoder frozen

• Loss: MSE

• Content/style balance: Balanced

• LoRA configuration:

  • Linear rank 32 / Conv rank 16

  • Alpha 32 / 16

Model Base

Optimized for the Z-Image model family, specifically:

• Base: ostris/Z-Image-De-Turbo

• Extras: Tongyi-MAI/Z-Image-Turbo

Trigger Word

Recommended LoRA strength: 1.0

midriff

Sampler Prompts

• "DJ at a night club, fisheye lens, smoke machine, laser lights, holding a drink, midriff"

• "Model showing off a new t-shirt at the beach, a shark jumping in the background, midriff"

• "Guitarist performing on stage, laser lights, punk rock scene, midriff"

• "Studio fashion photo, white background, medium shot, modeling clothing, midriff"