luca-voss

luca-voss (LoRA)

Trigger: [luca-voss] oai_citation:2‡config.yaml

A character LoRA trained to generate a consistent male model identity ("luca-voss") with photorealistic/editorial portrait aesthetics. The model was trained on Tongyi-MAI/Z-Image-Turbo with a LoRA network configuration optimized for identity consistency and realistic skin detail. oai_citation:3‡config.yaml

---

Base model

• Tongyi-MAI/Z-Image-Turbo oai_citation:4‡config.yaml

---

Training details

• Type: LoRA (Diffusers format)
• Steps: 3000
• Batch size: 1
• LR: 1e-4
• Noise scheduler: flowmatch
• Optimizer: adamw8bit
• Train UNet: true
• Train text encoder: false
• Precision: bf16
• Quantization: qfloat8 (model + text encoder)
• Resolutions used: 512 / 768 / 1024 oai_citation:5‡config.yaml oai_citation:6‡log.txt

LoRA network settings

• linear rank/alpha: 32 / 32
• conv rank/alpha: 16 / 16
• lokr_full_rank: true oai_citation:7‡config.yaml

---

How to use

Prompting

Use the trigger token at the start of your prompt:

Example (portrait/editorial): [luca-voss], 20-23-year-old man, dark brown strictly slicked-back hair, hazel-brown eyes, realistic skin texture, modern editorial portrait, photorealistic

Style / scenario examples

(These were used as sample prompts during training) oai_citation:8‡config.yaml

• Haute couture runway
• White tie (tailcoat / frack)
• Baywatch lifeguard vibe
• 1950s greaser / James Dean inspired
• Cyber-noir / Blade Runner inspired
• Cowboy / western vibe

Tip: For maximum identity stability, keep [luca-voss] at the beginning and add scene/outfit variations after.

---

Recommended generation notes

• If you see overly glossy or wet-looking skin artifacts, reduce your detailer/upscale denoise and reinforce matte/anti-shine descriptors in negative prompts.
• For stronger identity, prefer close-up / head-and-shoulders shots and consistent lighting across a series.

---

Limitations

• As with most character LoRAs, extreme stylization or very low-resolution generations can reduce identity consistency.
• Performance may vary across samplers and upscalers.

---

Training config (summary)

This LoRA was trained via a diffusion trainer job with the following key parameters (see config/log for full details). oai_citation:9‡config.yaml oai_citation:10‡log.txt