soft-anisotropic-diagrams

A kernels package providing CUDA + Metal ops for soft anisotropic Voronoi diagrams: rasterization, jump-flooding (JFA) initialization, gradient-based site optimization (Adam), densification, and pruning.

Install / load

from kernels import get_kernel

voronoi = get_kernel("blanchon/soft-anisotropic-diagrams")

# Now use the exposed ops:
img   = voronoi.render_voronoi(...)
sites = voronoi.init_candidates(...)
voronoi.run_jfa(...)

Exposed ops

Rendering & initialization

• render_voronoi — rasterize a (soft) anisotropic Voronoi diagram from a set of sites
• init_candidates — uniform candidate site initialization
• init_gradient_weighted — gradient-magnitude-weighted candidate initialization
• run_jfa — Jump Flooding Algorithm for nearest-site assignment
• precompute_site_data — per-site precomputation used by the renderer

Optimization

• compute_gradients_tiled — tiled gradient accumulation against a target image
• adam_update — fused Adam step on site parameters
• update_candidates — candidate refinement step
• radix_sort_pairs — GPU radix sort used by the pipeline

Densification / pruning

• compute_site_stats
• compute_densify_score_pairs
• compute_prune_score_pairs
• write_split_indices
• split_sites
• prune_sites

Metrics

• compute_psnr
• compute_ssim

Build variants

Prebuilt for: