Back to Projects 🎨 Generative AI & Fashion Open Source

Afro Fashion Stable Diffusion

A Stable Diffusion model fine-tuned on African fashion data to generate culturally relevant fashion images — part of the InFashAI initiative to create more inclusive and diverse AI models.

"Incorporating African fashion styles into cutting-edge AI technologies, helping stylists, designers, and creators in their creative process."

Standard generative models tend toward Western fashion styles. By fine-tuning Stable Diffusion on nearly 100,000 African fashion images from Afrikréa, we created a model that understands and generates culturally authentic African clothing — from wax prints to kente fabrics.

How It Works

From data curation to image generation in three steps.

1

Data Curation

Cleaned and standardized ~100K images from Afrikréa with prompts covering clothing type, material, fabric, color, and gender.

2

Fine-Tuning

Fine-tuned Stable Diffusion v1.4 on the curated dataset, producing 5 checkpoints at 20K, 40K, 60K, 80K, and final training steps.

3

Generation

Generate African fashion images from text prompts with culturally specific attributes like wax, ankara, kente, and vlisco fabrics.

What Fine-Tuning Improves

The fine-tuned model outperforms the base Stable Diffusion in areas specific to African fashion.

🪡

African Materials

Accurately represents wax, kente, ankara, and vlisco fabrics that the base model misses.

📿

Cultural Details

Captures accessories like cowries and traditional ornaments omitted by standard models.

👗

Clothing Styles

Generates proper mid-length cloths, boubous, and other African garment types from prompts.

🎨

Fabric Patterns

Produces authentic African patterns and prints rather than defaulting to Western styles.

Technical Details

Built on proven generative AI architecture with accessible hardware requirements.

🧠

Base Model

Stable Diffusion v1.4 from CompVis — a latent text-to-image diffusion model.

📊

Training Data

~100,000 African fashion images with curated text prompts from Afrikréa platform.

⚙️

Training

96K max training steps, 512px resolution, FP16 mixed precision, 30 GB VRAM.

🚀

Inference

Only 8 GB VRAM required. 5 checkpoints available for different quality trade-offs.

Stable Diffusion HuggingFace Diffusers PyTorch Accelerate Python CUDA

Partners

Made possible through collaboration with leading organizations in research and African fashion.