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.

Thanks for asking, C2PM isn’t meant to replace raster formats it serves as an intermediate format for workflows where users repeatedly operate on color groups (palette swaps, region recoloring, mask generation, etc). In a standard image, every color swap requires rescanning all pixels. In C2PM, the color to pixel index is already stored, so repeated swaps or multi-color operations become O(1) per color instead of O(P) per operation. It’s designed for tools where color groups matter more than spatial order, not for real-time display or GPU textures.