Brown-Hue-southasian-lora

<Gallery />

Note : Model limited to female portraits for now. Improved version is being worked on.

Brown-Hue (SDXL) — South Asian Realism LoRA

🔮 Model Overview

Brown-Hue is a high-quality LoRA built for Stable Diffusion XL (SDXL), designed to deliver authentic South Asian realism in AI-generated female portraits. This LoRA specializes in capturing rich, warm skin tones, expressive eyes, and soft, natural lighting—bringing a grounded and elegant look to your outputs.

Trained on a focused dataset of 60 high-resolution images, Brown-Hue enhances visual identity with cultural nuance and stunning detail. Whether you’re aiming for character design or photographic realism, this model excels at producing portraits that feel soulful and true-to-life.

---

✨ Features

• South Asian Realism – Designed to authentically represent South Asian features, tones, and expressions.
• Specialized in Female Portraits – Optimized for realistic depictions of women with cultural fidelity and aesthetic grace.
• High-Quality Dataset – Trained on a handpicked collection of 60 high-res images for maximum visual quality.
• Photorealistic Output – Ideal for photography-style compositions, editorial aesthetics, and character-driven artwork.

---

🏗️ Training Details

• Base Model: stabilityai/stable-diffusion-xl-base-1.0
• Dataset Size: 60 images (high-resolution, curated)
• GPU Setup: Dual NVIDIA T4 (2x T4)
• Precision: fp16
• LoRA Weight: 0.6 (recommended)
• Clip Skip Strength: 1
• Sampling Steps: 40–45
• CFG (Guidance Scale): 7.0
• Scheduler: DPM++ 2M SDE Karras

---

🛠️ Usage Instructions

from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler
from diffusers.utils import load_image
import torch

# --- SETTINGS ---
model_path = "stabilityai/stable-diffusion-xl-base-1.0"
lora_repo = "HyperX-Sentience/Brown-Hue-southasian-lora"
output_path = "output.png"

prompt = (
    "indo-realism, Golden Allure, A close-up shot of a young woman with caramel-toned skin and a delicate jawline. "
    "Her luminous black hair is styled sleek and straight. She wears an off-shoulder champagne-colored top with fine golden embroidery. "
    "Her subtle makeup emphasizes her sharp features. A softly blurred warm-toned background complements the golden hues. "
    "Shot on a Canon EOS R6, F/2.2, ISO 100 --style raw --ar 3:4 --chaos 12."
)
negative_prompt = "text, watermark"

# --- LOAD BASE PIPELINE ---
pipe = StableDiffusionXLPipeline.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    variant="fp16"
).to("cuda")

# Use DPM++ 2M SDE ~ DPMSolverMultistep
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

# --- LOAD LORA FROM HUB ---
pipe.load_lora_weights(lora_repo)
pipe.fuse_lora()

# --- GENERATE IMAGE ---
generator = torch.manual_seed(663477859063917)
image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=45,
    guidance_scale=7.0,
    generator=generator,
    width=1024,
    height=1024,
).images[0]

# --- SAVE ---
image.save(output_path)

---

🎯 Intended Use

• Portrait Photography Styles – Perfect for generating editorial, cinematic, and natural-looking portraits with South Asian aesthetics.
• Character Design – Useful for creating culturally rich, realistic female characters.
• Fashion & Concept Art – Enhances illustrations and concept scenes with grounded realism and expressive beauty.

---

⚠️ Limitations

• Female Portrait Focus – This LoRA is currently optimized only for generating female portraits.
• Prompt Sensitivity – Works best with well-crafted, descriptive prompts to guide tone, lighting, and detail.
• Style Bound – While flexible, Brown-Hue is realism-focused and may not suit abstract or stylized workflows.

---

📌 Notes

This LoRA is ideal for use with the SDXL base model and benefits from balanced sampling parameters. For consistent results:

• Stick within the recommended LoRA weight range (0.6)
• Use 40–45 inference steps with a CFG of 7.0
• Keep clip strength at 1.0
• Use the DPM++ 2M SDE Karras scheduler for optimal lighting and texture output

For best results, pair with prompts that emphasize natural lighting, emotion, and realism.