Model description
🧩 LoRA Name: Astrid_ZimageTurbo
> Base Model: Tongyi-MAI / Z-Image-Turbo
> LoRA Type: Character / Actress
> Trigger Word(s): Astrid
> Recommended Weight: 0.6 – 1.0 (ideal range ≈ 0.8–0.9)
✨ Overview
> This LoRA is trained to capture the distinctive identity, facial structure, and ethereal presence of Astrid Bergès-Frisbey, optimized specifically for Z-Image-Turbo. The model emphasizes likeness fidelity, subtle sensuality, and editorial realism, with a strong bias toward cinematic, European, and high-fashion aesthetics.
It excels at portraying Astrid’s delicate bone structure, intense gaze, and understated elegance across studio editorials, natural-light urban scenes, luxury interiors, and intimate portrait compositions.
🎨 What This LoRA Is Good At
> High likeness accuracy with preserved facial geometry
> Moody, cinematic, and editorial portraiture
> Natural skin texture with realistic micro-detail
> European fashion and lifestyle aesthetics
> Strong performance in low-key lighting, chiaroscuro, and soft daylight
> Portrait, half-body, and mirror-selfie compositions
⚙️ Usage Notes
> Use the trigger word “Astrid” once in the prompt
> Avoid stacking other character LoRAs to preserve identity fidelity
> Performs best with realistic lenses (50–85mm) and fashion/editorial lighting
> Guidance scale ≤ 1 recommended for native Z-Image-Turbo behavior
> Responds well to detailed wardrobe, makeup, and lighting descriptions
🧪 Example Prompt
Astrid, cinematic editorial portrait, soft smokey eyes,
natural skin texture, low-key studio lighting,
European fashion aesthetic, ultra-realistic
📐 Training Details
> Dataset: Captioned image dataset
> Training Resolutions: 512 / 768
> Total Steps: 3,500
> Batch Size: 1
> Content vs Style: Balanced
> Noise Scheduler: FlowMatch
> Optimizer: AdamW 8-bit
> Learning Rate: 1e-4
> Precision: BF16
> Text Encoder Training: Disabled
> UNet Training: Enabled
> Quantization: QFloat8 (model + text encoder)
> Architecture: ZImage Turbo LoRA (LoKR full-rank enabled)
⚠️ Limitations / Notes
> Not designed for anime or heavily stylized/cartoon outputs
> Extreme facial exaggeration or ultra-wide lenses may reduce likeness
> Best results achieved when paired with realistic photography prompts
Trigger words
You should use Astrid to trigger the image generation.
Download model
Download them in the Files & versions tab.
