LoRA: Dahlia Custom Character

Krea 2:
Activation Trigger Word: dahlia_char

Unlike previous versions that relied solely on natural language, this model introduces a hardcoded, unique trigger token to guarantee facial accuracy and isolate her identity from baseline concepts. May retrain to isolate face from outfit dataset bleed.

Trigger Word: None required — prompts naturally.
This LoRA was trained on 39 images with natural-language captions, designed to provide consistent generations of a fit, athletic woman across a wide range of settings. Example images were generated with Flux-Krea -- Trained on Flux-Dev, so likeness will be more preserved and consistent on Dev.

SDXL model is best used with DMD2 Lora -- better outputs, character consistency, prompt adherence. Model was probably overfit, but the DMD lora resolves the issue.

📝 Training Details

The dataset focused on portraits with expressive emotions, varying fashion styles, and natural lighting, so the LoRA adapts well to most realistic scenarios. Dataset created with Qwen-Edit from a single image generation made with BigLovePonyXL to test the versatility of SOTA image editing models.

🎨 How to Use

• Works best with portrait photography, cinematic realism, or fashion concepts.

• No trigger word needed. Simply describe the subject in natural prompts (e.g., “athletic woman, portrait, natural lighting, expressive smile”).

• Strong activation with tags like: athletic, portrait, realistic, expressive face, fashion.

✅ Recommended Settings

• Weight: 0.6–0.9 for subtle consistency, 1.0–1.2 for stronger adherence.

• Model Base: Any Flux or SDXL (with DMD2 lora) model tuned for realism or portrait photography.