Ccadic/erwan-ZImage-Turbo-LoRA

Erwan — Z-Image-Turbo Character LoRA

A character identity LoRA trained on Z-Image-Turbo using Ostris AI-Toolkit.

Overview

Erwan is an original fictional male character, approximately 27 years old, with Mediterranean features, dark curly hair, brown eyes and a youthful cinematic appearance.

The goal of this LoRA is to maintain a recognizable character identity across different environments, lighting conditions, clothing, camera angles and photographic compositions.

Model information

Parameter	Value
Base model	Tongyi-MAI/Z-Image-Turbo
Training software	Ostris AI-Toolkit
Training images	35
Training steps	2,000
LoRA rank	16
LoRA alpha	16
Optimizer	AdamW 8-bit
Learning rate	1e-4
Training precision	BF16
Training resolutions	512 / 768 / 1024
Trigger word	erwan_id
Format	Safetensors

Training used the Z-Image-Turbo training adapter v2.

Usage

Load the original Z-Image-Turbo model and apply the character LoRA.

Trigger word: erwan_id

Example prompt:

Cinematic photograph of erwan_id man standing in a futuristic spacecraft, realistic facial features, natural expression, dramatic lighting, professional photography

Suggested LoRA strengths to evaluate: 0.6, 0.8, 1.0 and 1.2.

These are experimental values, not validated optimal settings.

Example prompts

Studio portrait

Professional studio photograph of erwan_id man, neutral gray background, soft lighting, realistic skin texture

Concert

Cinematic photograph of erwan_id man playing acoustic guitar on stage, colorful concert lighting, natural expression

Adventure

Photorealistic image of erwan_id man exploring an ancient archaeological site, desert landscape, cinematic lighting

Training notes

The dataset contains 35 character images with varied camera angles, environments, poses, lighting conditions and compositions.

The model was trained using a quantized Z-Image-Turbo backbone with a training adapter.

Training checkpoints were saved every 250 steps. The published model corresponds to the final 2,000-step checkpoint.

Limitations

• Character identity may vary with extreme camera angles or complex compositions.
• Exact facial consistency is not guaranteed.
• Text rendering and hand anatomy depend partly on the base model.
• Further quantitative identity evaluation is recommended.

Credits

Created by Philippe Cadic.

Base model: Tongyi-MAI/Z-Image-Turbo.

Training framework: Ostris AI-Toolkit.