🧩 3D LoRA Low-Rank Adapter: Eckart-Young Rank Limit, Gradient Descent on B and A, and Old-Task Drift
Three Blender-built frames hold 16 x 16 bar fields: the update a task wants, the rank-r adapter s B A trained by gradient descent from B = 0, and the residual, whose energy settles on the Eckart-Young spectrum tail; old-task drift is measured against a full fine-tune and six independent checks run on load.
⚙ Under the hood
Three Blender-built frames hold 16 x 16 bar fields: the update a task wants, the rank-r adapter s B A trained by gradient descent from B = 0, and the residual, whose energy settles on the Eckart-Young spectrum tail; old-task drift is measured against a full fine-tune and six independent checks run on load.
LoRALow-Rank AdaptationFine-TuningEckart-YoungSingular ValuesGradient DescentCatastrophic ForgettingParameter-Efficient
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install