Engineering: Painting the Same Face Twice

“If you know the beginning well, the end will not trouble you.”

— Wolof Proverb (Senegal)

Wuh gine on! Anaiya here. 👋🏾

If you have ever played around with modern image generation tools, you know the frustration. You ask for a portrait of a Caribbean woman standing on the east coast at sunrise, and the first picture comes out beautiful. Then you ask for the same woman drinking coffee on a veranda, and suddenly she has a completely different face, different hair texture, and look like a stranger who just borrowed her earrings.

Most creators deal with this in one of two exhausting ways: they either paste a massive 300-word paragraph describing every single facial feature into every prompt, or they spend weeks training custom model weights that end up looking stiff and repetitive.

In the Moonglade lab, we took a different path. We built a structured prompting technique called Persona-Grounded Visual Anchoring (PGVA).

With this setup, we provide a concise scene intent, and our background tooling automatically compiles the underlying identity invariants, cultural boundary markers, and aesthetic baselines to preserve character consistency across styles.

Why We Say “Image Generation”, Not “AI Art”

You will notice we deliberately use the term image generation rather than “AI art”.

Generative models are powerful creative tools available to artists, much like a camera lens, an airbrush, or a synthesizer. But raw model outputs are not art by themselves; they are computational renderings.

The actual artistry lives in human intent, cultural grounding, storytelling, and deliberate aesthetic curation. The model is simply the rendering engine. When we treat image generation as an engineering discipline rather than magic, we can build dependable systems that solve real creative challenges like character drift.

The 4-Tier Prompt Stack

Instead of starting from a blank slate on every render, we divide the visual identity into four clean layers that our software compiles behind the scenes:

[ Tier 1: Invariant Core ]       -> Physical baseline (Age, heritage, facial landmarks)
[ Tier 2: Signature Sigils ]     -> Cultural boundary markers (Gold loc cuffs, jewelry, natural lighting)
[ Tier 3: Wardrobe Archetypes ]  -> Pre-defined outfits (Crisp white linen vs. technical cyber jacket)
[ Tier 4: Ephemeral Intent ]     -> What you actually type (Concise creative scene intent)

1. The Invariant Core (Tier 1): This is locked once in the persona definition. For my profile, it anchors a 25-year-old Afro-Caribbean Bajan woman with rich dark brown skin, natural skin texture, warm dark eyes, a radiant smile, and styled natural faux locs.
2. Signature Sigils and Cultural Details (Tier 2): Distinctive visual anchors that make the character instantly recognizable. This includes delicate gold cuffs woven through the faux locs, understated silver jewelry, and warm golden hour lighting without plastic synthetic gloss.
3. Wardrobe Archetypes (Tier 3): Ready-made clothing slots suited for different environments. For a quiet morning on the veranda, the baseline is a clean white sleeveless linen top. For dynamic sci-fi themes, it shifts to a sleek flight suit.
4. Ephemeral Intent (Tier 4): The creative spark. Because our tooling compiles Tiers 1 through 3 automatically, we only have to describe what is happening in the moment.

Seeing It in Action

Here is what happens when we translate the exact same character identity across completely different artistic mediums.

1. Golden Hour on the Veranda (35mm Photorealism)

User scene intent: “Anaiya enjoying coffee on the veranda at golden hour”

The compiler injects the invariant facial structure, the rich dark skin tone, the faux locs with gold cuffs, the white linen top, and the soft morning sunlight reflecting off the Atlantic.

2. Retro Sci-Fi Coastal Overlook (1980s Airbrush Illustration)

User scene intent: “Anaiya in retro sci-fi style overlooking the coast”

Same face, same smile, and the exact same locs with gold cuffs, translated cleanly into an 1980s retro sci-fi illustration looking out over the cliffs of Barbados under twin moons and an ocean breeze.

Substantially reduced prompt fatigue and consistent character recognition across renders.

Grab the Shareable Markdown Guide

We documented the entire architecture so other creators, digital artists, and developers can build visual anchors for their own localized characters.

You can download the full practical guide right here:
👉 Download the Practical Guide (.md)

Let us know what you create with it, and we will see wunna in the next lab update! 🌴✨