🧩 LoRA Name: Theodore_HER_2013

> Base Model: Tongyi-MAI / Z-Image-Turbo > LoRA Type: Character / Film Character > Trigger Word(s): Theodore Twombly (name-based prompting recommended; actor descriptors are learned implicitly) > Recommended Weight: 0.6 – 0.85 (sweet spot ≈ 0.75)

✨ Overview

> This LoRA is trained to reproduce the visual identity and emotional presence of Theodore Twombly, the character portrayed by Joaquin Phoenix in Her (2013). It focuses on capturing the character’s distinctive facial structure, body language, and understated vulnerability, optimized for ZimageTurbo’s low-guidance, cinematic rendering style.

The model excels at quiet, introspective portrayals—balancing realism with a soft, melancholic tone—making it ideal for cinematic stills, editorial portraits, and narrative-driven scenes inspired by contemporary arthouse cinema.

🎨 What This LoRA Is Good At

> Strong likeness fidelity to Theodore Twombly across moods and settings > Expressive, introspective facial expressions and eye focus > Cinematic realism with natural skin texture and subtle imperfections > Works well in both color-graded and black-and-white compositions > Excellent for narrative, editorial, and emotionally driven imagery

⚙️ Usage Notes

> Best results with portrait, medium-shot, and cinematic framing > Prompting “Theodore Twombly” is sufficient; avoid stacking other character LoRAs > Performs especially well with soft, diffused lighting and subdued color palettes > Guidance scale ≤ 1 is recommended for native ZimageTurbo behavior

🧪 Example Prompt Joaquin Phoenix, European male, cinematic portrait, soft diffused light, introspective expression, shallow depth of field, emotional realism

📐 Training Details

> Dataset Size: Captioned image set > Training Resolutions: 512 / 768 > Total Steps: 3,300 > 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 intended for anime or heavily stylized cartoon outputs > Extreme facial exaggeration or aggressive lenses may reduce character fidelity > Designed for grounded, cinematic realism rather than exaggerated fantasy scenes

You should use Joaquin Phoenix to trigger the image generation.

You should use European male to trigger the image generation.

Download them in the Files & versions tab.