South-Indian-Beauty - HiDream

Trigger Word: south-indian-beauty

Base Architecture: Hidream O

Training Depth: 5 Epochs| 14 Repeats | 2,013 Steps

This model shifts your fine-tune directly into the HiDream architecture, taking full advantage of its powerful Diffusion Transformer (DiT) backbone. Unlike standard pipelines that struggle with localized details, HiDream integrates an incredibly robust semantic text engine. This makes it the ideal system for learning complex, highly specific cultural aesthetics—such as the unique bone structures, authentic skin textures, and rich melanin palettes that define South Indian beauty.

By locking in a configuration of 5 full epochs with a precise Unet LR of 0.0001, this LoRA avoids the "washed-out" skin tones common in lesser models, anchoring a vibrant and mesmerizing character consistency.

This configuration is tuned to perfectly match HiDream's high-fidelity inference pipeline:

• Epochs: 5 (Optimized training duration for stable weight distribution)

• Total Steps: 2,013

• Num Repeats: 14 (Balanced reinforcement per asset to prevent overfitting)

• Learning Rate: 0.0001 (Slow, high-accuracy learning to preserve natural skin grain)

• Resolution: 1024 x 1024 (Native square high-res bucketing enabled)

• Clip Skip: 1 (Maintains native text-conditioning precision)

For optimal rendering on HiDream-based pipelines:

• LoRA Scale/Weight: 0.8 - 0.95

• Inference Steps: 28 - 40

• CFG Scale: 1.5 - 3.0

• Resolution: Best generated at 1024x1024 or higher aspect ratios.

This 5-epoch HiDream variant is built for those who appreciate true realism and cultural accuracy without over-baking the base weights. The combination of 14 repeats over 2,013 steps strikes the absolute perfect balance for flexible prompting. Please share your generations and post your feedback below—your creativity is the absolute best reward!