metadata
license: creativeml-openrail-m
base_model: black-forest-labs/FLUX.1-dev
tags:
- lora
- flux
- ai-toolkit
- character-style
- person
- z-image-turbo
datasets:
- simone-v2
Model Card: Simone-ZIT-v2.2 (Z-Image De-Turbo)
This is a specialized character LoRA optimized for Z-Image De-Turbo (De-Distilled). Like the previous versions, it is trained as a Style, meaning the character is embedded into the model's weights without a specific trigger word.
Training Philosophy
The model focuses on high-fidelity character consistency across various environments while maintaining the speed and quality benefits of the Z-Image Turbo architecture.
Caption Strategy:
- No Trigger Word: The character's name and defining features (hair/eye color) were omitted from captions to bake them directly into the subject weight.
- Full Environment Decoupling: Meticulous tagging of clothing, backgrounds, and lighting ensures the character remains a constant subject while the scene remains flexible.
- Flexible Generation: The character manifests automatically when describing a "woman" or "subject."
Usage & Prompting
Describe a woman and her surroundings. The model will automatically apply the Simone character style.
- Example Prompt:
A woman wearing a silk blouse and tailored trousers, sitting in a modern sunlit cafe. - LoRA Strength: Recommended 0.7 - 1.0 for character accuracy.
- Compatibility: Best used with Z-Image Turbo/De-Turbo workflows.
Technical Specifications
| Parameter | Value |
|---|---|
| Trigger Word | None (Trained as Style) |
| Model Architecture | Z-Image De-Turbo (De-Distilled) |
| Rank (Dimension) | 128 |
| Batch Size | 4 (Constant throughout training) |
| Precision | float8 |
| Total Steps | 4000 |
| Save Frequency | Every 200 steps |
| Training Hardware | NVIDIA H200 Tensor Core GPU |
| Training Toolkit | Ostris - AI Toolkit |
Model card generated for the Simone-ZIT series.