Balledk/cowl-ZIT-v1.0

Model Card: Flux LoRA - cowl (Z-Image De-Turbo)

This is a LoRA model trained on a curated dataset of high-quality images (1024x1024) using the Ostris - AI Toolkit.

Architecture: Specifically optimized for Z-Image De-Turbo (De-Distilled).

Training Settings

Parameter	Value
Trigger Word	cowl neck
Model Architecture	Z-Image De-Turbo (De-Distilled)
Batch Size	2
Rank (Dimension)	48
Precision	float8 (Transformer & Text Encoder)
Save Frequency	Every 200 steps
Total Training Steps	4000

Prompting Strategy & "Mix & Match"

This LoRA was trained using granular, tag-based descriptive captions. This approach "de-couples" specific attributes, allowing you to freely combine features from across the dataset.

• Modular Control: You are not limited to the training images. You can mix attributes—for example, taking a silk texture from one style, a deep drape from another, and setting the color to emerald green.
• Tag-Based Precision: Since every detail was tagged, the model performs best when you describe the specific materials (satin, jersey, wool), finishes, and colors you want to see.

Example Prompt:

cowl neck, emerald green color, satin material, deep draped folds, elegant aesthetic, studio lighting

Usage

• Inference: Use settings compatible with Z-Image Turbo/De-Turbo.
• LoRA Strength: 0.6 - 1.0 (Start at 0.8).
• Precision: Trained in float8. Ensure your environment (ComfyUI, Diffusers, etc.) supports this for optimal results.

Training Infrastructure

• Toolkit: Ostris AI Toolkit
• Dataset Size: [Indsæt antal] images (1024x1024)
• Training Method: LoRA

---

Model card generated for the cowl series.