Trigger Word: south-indian-beauty
Ecosystem: ERNIE (Enhanced Representation through Knowledge Integration)
Training Engine: AI-Toolkit
Dataset: 85 High-Resolution Onsite Generations
Rank / Alpha: 32 / 16
Repeats: 22
Total Steps: 3,117
LR Scheduler: Cosine with Restarts
Noise Offset: 0.1
Text Encoder LR: 0.00005
Unet LR: 0.0005
Epochs: 10
Resolution: 1024
Min SNR Gamma: 5
Optimizer: AdamW 8-bit
Sample prompts:
south-indian-beauty, sophisticated mature woman, wearing a traditional heavy silk Kanchipuram saree with metallic zari, standing in a historic temple hallway, flickering oil lamps creating dramatic shadows, deep brown eyes, intricate antique gold temple jewelry, 85mm f/1.8, hyper-realistic
south-indian-beauty, stunning mature woman with an elegant aura, wearing a tailored navy blazer over a crisp floral saree, standing in a contemporary glass-walled gallery, soft afternoon lighting, sharp focus on expressive brown eyes, realistic skin maturity, vogue editorial style.
south-indian-beauty, beautiful mature woman, looking out at a rainy garden through a window, soft natural morning light, wearing a simple hand-loomed cotton kurta, salt-and-pepper hair in a loose bun, focus on realistic skin pores and warm amber depth in her eyes, peaceful atmosphere, cinematic photography.
Description
Update: This version utilizes the same high-quality dataset as the previous iteration but features optimized training parameters to improve generalization. By reducing the Epochs to 10 and the total Steps to 2,125, this version aims to strike a better balance between subject accuracy and prompt flexibility, reducing the "fried" or over-baked look common in higher-step LoRAs.