Building Nalybecks Afrofuturistic AI: A Quiet MVP, Computer Vision, and the Road to Generative Models

For a long time, I’ve been building quietly.

Not because nothing was happening — but because I was experimenting, iterating, and learning before feeling ready to share anything publicly. This post is part of that shift: documenting the process, not just the outcome.

The idea

Nalybecks Afrofuturistic AI started as a question:

What would beauty-tech look like if it were designed with African hair, culture, and aesthetics at the center — not as an afterthought?

Most existing hair try-on or beauty AI systems struggle with African hair textures, volume, braids, twists, and knots. That gap isn’t accidental; it’s a data, representation, and modeling problem.

Before jumping straight into heavy generative models, I wanted to understand the problem space end-to-end.

That’s where the MVP comes in.

What the MVP is (and why it exists)

This project is currently at the MVP (Minimum Viable Product) stage.

Here, MVP doesn’t mean “finished” or “polished.” It means the smallest working system that validates the pipeline, surfaces constraints, and guides future research decisions.

The app is a web-based MVP, deployed on Streamlit (https://nalybecks-afrofuturistic-ai-9wujmg77e7xxajhhvcvvbr.streamlit.app/) and accessible on Android and iOS via the browser. This choice allows for fast iteration and learning without prematurely committing to native platforms.

Technical overview of the current system

1. Input and preprocessing

Users upload a selfie, which is processed using OpenCV and PIL for basic image handling and normalization.

2. Face analysis

The system uses MediaPipe Face Mesh to detect facial landmarks. From these landmarks, it estimates:

• Face width
• Face height
• Aspect ratio

These features are intentionally simple but interpretable.

3. Hairstyle recommendation

A lightweight heuristic-based logic maps face geometry to hairstyle categories (e.g., twists, cornrows, afro crown, galactic knots).

This layer is:

• Fast
• Explainable
• Suitable for an early MVP

4. Hairstyle try-on

The current try-on experience uses:

• Pre-designed transparent PNG hairstyles
• Dynamic resizing relative to detected face width
• Alpha blending via OpenCV for compositing

This creates a basic visual approximation of a hairstyle on the user’s image.

Known limitations (by design)

The limitations of the current MVP are expected and intentional:

• Overlays are 2D
• No modeling of hair volume or curl dynamics
• No adaptation to lighting or head pose
• Limited identity preservation
• No photorealistic synthesis

These are not bugs — they are the natural ceiling of classical computer vision overlays.

In many ways, the MVP’s job is to make these limitations obvious.

Why not jump straight to GANs or Diffusion?

Because generative models are only as good as:

• The problem definition
• The data
• The evaluation criteria

Building this MVP first helped clarify:

• Where classical CV breaks down
• What “realism” actually means in this context
• Which parts of the pipeline need to be generative vs. geometric

This groundwork is essential before moving into GANs and Diffusion models.

The next phase: Generative models

The next stage of this work will explore:

• Diffusion models for photorealistic hair synthesis
• GAN-based approaches for identity-preserving generation
• Style-conditioned generation focused on:
• African hair textures
• Braids, twists, knots, and volume
• Cultural and aesthetic accuracy

This also raises deeper research questions around:

• Dataset representation
• Bias and generalization
• What “realistic” means across cultures

Why this matters (especially from an African context)

Beauty-tech AI is not neutral.

When African hair and aesthetics are underrepresented in training data, the resulting systems fail — not because the problem is unsolvable, but because it hasn’t been centered.

This project is a small, early attempt to think differently:

• From African-first design
• From Ugandan innovation
• From a research-oriented, learning-driven perspective

Sharing the journey

I’m sharing this work now not because it’s finished, but because progress deserves documentation.

Innovation often looks like quiet iterations long before it looks like a polished product.

As this project evolves, I plan to keep sharing:

• Technical decisions
• Research challenges
• What works — and what doesn’t

The journey from classical CV to generative models is ongoing.

If you’re working in applied AI, computer vision, or African-centered technology, I’d love to exchange ideas.

More to come.

ArtificialIntelligence #ComputerVision #MachineLearning #AfricanInnovation #BeautyTech #WomenInTech