ARD-KMeans++ modifies KMeans++ initialization by incorporating local density estimates alongside distance. This reduces centroid bias toward sparse regions in skewed or multimodal distributions and improves convergence stability without changing the K-Means objective. Early experiments show fewer degenerate initializations and lower variance across runs compared to standard KMeans++, particularly on non-uniform datasets. Critical feedback and failure cases welcome.