What it was.
A lot of avatar generation looks convincing from the front and falls apart as soon as the head turns. That makes it fine for a still image and weak for realtime use.
I approached the problem as a workflow instead of a single model pass: combine multiple viewpoints, clean the result in Blender, and push toward output that could survive use inside an actual product.
That added work compared with one-click generation, but it improved consistency, embodiment, and runtime viability.
The output fed directly into later agent and avatar work like Jessica, so it functioned as enabling product R&D rather than a disconnected art exercise.
The pipeline combined five-photo input, multi-view synthesis, Blender cleanup, and Unity-ready output that fed directly into later avatar work. It turned unstable generation into something much closer to production craft.
