BBB25 Digital Twin

Dozens of hours of video and audio were captured with Vitória, then used to train a digital twin that could answer authentically across a wide range of situations, not just recite lines.

Three techniques carried it: diffusion models for the image, voice cloning for the sound, and careful fine-tuning to keep her mannerisms intact. The goal was never a clone. It was to preserve her essence in every response, which is a much harder brief and the only one worth accepting.

With a curated dataset and generative automation, the first working version of the twin was delivered in ten days.

Vitória sat for it. Dozens of hours of video, audio and old interviews went into a curated training set, then diffusion models and voice cloning were fine-tuned on top of it until the double held up at phone distance.

The harder part was not the face. It was the judgement. The twin had to answer a live show it was not watching, in her register, without ever sounding like a brand. Being close to Vitória is what made that possible, and it is why the work reads as an extension of her rather than a costume.

Ten days from start to a first working version. Months of material after that, pre-built and held ready so it could be cut to whatever the house did that night.