sd-lora-nigerian-adire

Stable Diffusion LoRA: Nigerian Adire Style

<div align="center"> <img src="images/header.jpg" alt="Generated Examples" width="800"/> </div>

Model Description

This is a LoRA (Low-Rank Adaptation) model fine-tuned on traditional Nigerian adire textile patterns. It teaches Stable Diffusion 1.5 to generate images in the distinctive adire style—characterized by intricate geometric patterns, indigo blue coloring, and traditional Yoruba design elements.

Trigger phrase: nigerian_adire_style

Key Features

• Generates authentic adire-style patterns
• Lightweight: 50MB (vs 7GB full model)
• Fast inference: ~12 seconds per image
• Compatible with standard SD 1.5 workflows
• Built with production MLOps pipeline

---

Usage

Quick Start

from diffusers import StableDiffusionPipeline
import torch

# Load base model
pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16
)

# Load LoRA weights
pipe.unet.load_attn_procs("AfroLogicInsect/sd-lora-nigerian-adire")
pipe = pipe.to("cuda")

# Generate
prompt = "a nigerian_adire_style painting of a sunset over Lagos"
image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
image.save("output.png")

Advanced Usage

# With custom parameters
image = pipe(
    prompt="a nigerian_adire_style portrait with geometric patterns",
    negative_prompt="blurry, low quality, distorted",
    num_inference_steps=75,
    guidance_scale=9.0,
    height=512,
    width=512,
    generator=torch.manual_seed(42)
).images[0]

With Diffusers Pipeline

from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16
)
pipe.unet.load_attn_procs("AfroLogicInsect/sd-lora-nigerian-adire")
pipe.to("cuda")

---

Training Details

Training Data

• Dataset: 30 high-quality images of traditional Nigerian adire patterns
• Resolution: 512×512 pixels
• Source: Curated collection of authentic adire textiles
• Preprocessing: Center-cropped and resized

Training Configuration

Base Model: runwayml/stable-diffusion-v1-5
Method: LoRA (Low-Rank Adaptation)
Rank: 4
Alpha: 4

Hyperparameters:
  Learning Rate: 1e-4
  Scheduler: constant
  Training Steps: 800
  Batch Size: 1
  Gradient Accumulation: 4
  Optimizer: AdamW (8-bit)
  Mixed Precision: fp16

Hardware:
  GPU: Tesla T4 (Google Colab)
  VRAM: 15GB
  Training Time: ~75 minutes

Training Process

The model was trained using DreamBooth with LoRA adapters:

1. Fine-tuning: Only attention layers were adapted (0.11% of total parameters)
2. Validation: Generated samples every 100 steps
3. Loss: Final loss converged to 0.042
4. Quality: CLIP score of 0.85 on validation set

MLOps Pipeline

This model was built with a complete production pipeline:

• Automated evaluation: Quality, speed, reliability checks
• Quality gates: Only promoted after passing thresholds
• Version control: Tracked with MLflow
• CI/CD: Automated testing and deployment
• Monitoring: Real-time inference tracking

---

Model Performance

Evaluation Metrics

Benchmark (on Tesla T4)

Resolution: 512×512 Steps: 50 Guidance: 7.5 Average Time: 12.3s P95 Latency: 14.1s Throughput: ~5 images/minute

---

Prompt Engineering Tips

Best Practices

# ✓ Good prompts (specific, descriptive)
"a nigerian_adire_style painting of Lagos cityscape at golden hour"
"a nigerian_adire_style portrait of a woman with geometric patterns"
"a nigerian_adire_style pattern featuring indigo and white colors"

# ✗ Poor prompts (too vague)
"adire style image"
"nigerian art"
"blue pattern"

Recommended Settings

num_inference_steps = 50-75  # Higher = better quality
guidance_scale = 7.5-9.0     # Higher = stronger style adherence
negative_prompt = "blurry, low quality, distorted, western style"

Advanced Techniques

Style Mixing:

prompt = "a nigerian_adire_style painting with Art Nouveau influences"

Subject Focus:

prompt = "a nigerian_adire_style portrait of [subject], detailed geometric patterns"

---

Limitations & Biases

Known Limitations

1. Style Consistency: May occasionally generate generic patterns instead of authentic adire
2. Color Range: Heavily biased toward indigo/blue (true to training data)
3. Resolution: Optimized for 512×512; higher resolutions may show artifacts
4. Text Rendering: Cannot generate readable text in adire style

Ethical Considerations

• Cultural Sensitivity: This model represents traditional Nigerian art. Use respectfully.
• Attribution: Generated images should not be misrepresented as authentic adire.
• Commercial Use: Review the CreativeML Open RAIL-M license for commercial applications.

Bias Statement

The model reflects the patterns present in the training data—traditional adire designs from specific Nigerian communities. It may not represent all regional variations or modern interpretations of adire art.

---

License

This model is released under the CreativeML Open RAIL-M license.

Key Points:

• You can use this model for personal and commercial purposes
• You can modify and redistribute the model
• You can NOT use it for illegal purposes
• You can NOT claim you created the base Stable Diffusion model

---

Citation

If you use this model in your research or project, please cite:

@misc{nigerian-adire-lora-2025,
  author = {Akan Daniel},
  title = {Stable Diffusion LoRA: Nigerian Adire Style},
  year = {2025},
  publisher = {HuggingFace},
  journal = {HuggingFace Model Hub},
  howpublished = {\url{https://huggingface.co/AfroLogicInsect/sd-lora-nigerian-adire}}
}

Blog Series:

@article{mlops-blog-series-2025,
  author = {Akan Daniel},
  title = {Building Production AI: A Three-Part MLOps Journey},
  year = {2025},
  url = {https://dev.to/afrologicinsect/building-production-ai-a-three-part-mlops-journey-38a8}
}

---

Resources

Related Links

• Full Tutorial (3-part blog series)
• GitHub Repository)
• Try the Demo (Gradio Space)
• YouTube Tutorial

Technical Documentation

• Diffusers Documentation
• LoRA Paper (arXiv)
• Stable Diffusion

Community

• Report issues: GitHub Issues
• Discussions: HuggingFace Discussions

---

Examples

<div align="center">

Example 1: Sunset Scene

<img src="images/example_1.png" width="400"/>

Prompt: a nigerian_adire_style painting of Lagos cityscape at sunset

Example 2: Portrait

<img src="images/example_2.png" width="400"/>

Prompt: a nigerian_adire_style portrait of a woman wearing traditional attire

Example 3: Pattern

<img src="images/example_3.png" width="400"/>

Prompt: a nigerian_adire_style pattern with indigo geometric shapes

</div>

---

Acknowledgments

• Base Model: Stable Diffusion 1.5 by RunwayML
• Training Method: LoRA by Microsoft Research
• Infrastructure: Google Colab, HuggingFace Hub
• Cultural Heritage: Traditional Nigerian adire artisans

---

Version History

v1.0.0 (2025-01-XX)

• Initial release
• Trained on 30 adire pattern images
• 800 training steps
• CLIP score: 0.85

---

<div align="center">

Built with ❤️ for cultural preservation through AI

Demo • Code • Tutorial

</div>