How We Built Maya: Training a Consistent AI Character

Maya is Vixn’s flagship creator — the face on our homepage, the same person in every one of the hundreds of images we’ve generated of her. Getting her there took multiple full identity trainings, one complete failure, and a lesson about realism we now enforce platform-wide. This is the honest build log.

The starting problem: a type, not a person

Early Maya was what every generator gives you: a beautiful girl who was a different beautiful girl every time. Same prompt, new face. The generator had learned “attractive woman matching this description” — a distribution of people, when what we needed was exactly one.

The fix is identity training: teaching a private model what this specific person looks like — her eye spacing, eyelid shape, nose width, lip ratio, jaw contour — so the distribution collapses to a single individual.

What the failed version taught us

Our second-generation Maya held identity well but had what we came to call beauty-ad skin: airbrushed, poreless, lit like a cosmetics campaign. Audiences read it as AI instantly. So for the next version we swung hard the other way — training her on heavily texture-rich, candid imagery.

It failed. The realism push dragged skin blemishes into her identity itself — and once damage is baked into a trained character, no prompt can cleanly remove it. Worse, suppressing it at generation time degraded her face. We threw the entire training away.

The lesson became a platform rule we now call pretty-first realism: camera-real texture — pores, natural oil sheen, real light behavior — belongs in a creator’s identity. Skin damage doesn’t. They are different physical things, and the dataset has to be curated to include one and exclude the other.

The brutal cull

The version of Maya you see today was trained on a ruthlessly curated set. Every candidate image was screened twice before a human ever ranked it: once for skin damage (any blemish auto-rejects, no matter how good the image), once for identity against her canonical face.

Then came the human pass — scoring each survivor as keep-plus, keep, or reject, with notes. Images that read “younger, softer, rounder” or collapsed into “generic Instagram beauty face” were cut even when they were gorgeous, because they were a type, not her. We’d rather train on a smaller set of brutal keepers than a large set of maybes — the maybes are what silently poison consistency.

The two strongest images — the ones that agreed most precisely on her facial geometry — were anchored in training with extra weight. They became her canon.

Stress-testing identity

Soft window light is forgiving; golden hour is not. Harsh, directional sunlight exposes identity errors faster than anything else, so poolside-at-sunset became our standard gauntlet. A Maya that survives golden hour with her face intact is a Maya that holds anywhere.

Every render is also scored automatically against her canon — identity, body, texture, and skin-damage checks — before it ever reaches a gallery. Drift gets caught and re-generated, not published.

What this means if you’re building a creator

Three takeaways from Maya that apply to any character you train on Vixn:

1. Curate brutally.Ten images that are unmistakably your character beat thirty that are “close.”
2. Texture yes, damage no.Real skin has pores and sheen. It doesn’t need blemishes to read as real.
3. Test in hard light. If identity survives golden hour, it survives your whole content calendar.

Maya’s full public set is at vixn.app/v/maya. Every image is the same person. That’s the product.