Home▸Machine Learning & Neural Networks▸LoRA Adapters

PEFT / LoRA Rank Bottleneck — 3D

3D visualization of LoRA's rank bottleneck: a wide grid of frozen weight cells (W) with a thin low-rank funnel (A/B, ΔW = B·A) squeezing a forward pass through r dimensions. Tune rank r, matrix dimension d and scale α and watch the funnel and parameter counts change live.

Machine Learning & Neural Networks3DEasy60 FPS📱 Mobile-adapted⇄ 2D version

This scene renders LoRA's low-rank adaptation spatially rather than as a matrix heatmap: a wide, dimmed platform of instanced cubes represents the frozen weight matrix W — untouched throughout training — while a narrow funnel above it shows the trainable path a forward pass actually takes through A and B, squeezing from d channels down to just r at the waist before expanding back out. Glowing pulses travel through the funnel to make the bottleneck tangible: widen the rank r and the waist opens up; shrink it and the funnel pinches to almost nothing while the frozen platform (d) stays exactly as large as before. The same trainable-parameter, percentage and compression stats as the 2D view update live as you drag the sliders.

⚙ Under the hood

Visualize the process with a 3D model where a wide platform of frozen-weight cubes narrows through an hourglass funnel, controlled by the rank r.

Machine LearningNeural NetworksLow-Rank Approximation

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install