For Z-Image Turbo

LoRA for Eve like body to bypass flat body of native ZiT especially when generating Asian.

This LoRA is specifically designed for the Z-Image Turbo (ZiT) architecture. It focuses on [Insert specific body type here, e.g., "an athletic and curvy silhouette"] while maintaining anatomical stability across different poses and outfits. Trained at $1024 \times 1024$ resolution, it excels at preserving skin textures and natural body proportions.

Quick Start Settings

• Trigger Word: EV3_ZiT

• Base Model: Z-Image Turbo (SDXL Distilled)

• Recommended Weight: 0.7 - 0.9 (Start at 0.8 for best results)

• Sampling Steps: 8 - 10 steps

• CFG Scale: 1.5 - 2.0 (Do not exceed 2.5 to avoid artifacts)

• Sampler: DPM++ SDE or Euler a

Prompting Tips

• Activation: Place the trigger word EV3_ZiT at the very beginning of your prompt.

• Clothing Flexibility: This LoRA was trained to be "clothing-aware." It works well with everything from swimwear and gym clothes to loose dresses and formal suits.

• Body Control: If you want to exaggerate the effect, use weight tagging in your prompt: (EV3_ZiT:1.2).

• Skin Details: For hyper-realistic skin, add keywords like highly detailed skin texture, natural skin pores, or subtle highlights.1

Recommended Negative Prompt

Since ZiT is high-speed, a minimal negative prompt is usually best:

(deformed iris, deformed pupils), text, watermark, blurry, low quality, bad anatomy, bad proportions, disfigured, (bad hands, extra fingers), metallic skin, high contrast.

Known Limitations

• Weight Sensitivity: At weights above 1.2, you may notice "color bleeding" or increased contrast, which is a trait of distilled Turbo models.

• Best Epoch: This version is based on Epoch 7, which provides the best balance between shape-accuracy and prompt flexibility.