Sliced Wasserstein embeddings + hand-rolled HNSW power color-scheme search over 48k paintings
Potential-Barber8658 · reddit · 2026-10-11
Palette Atlas searches 48,695 public-domain paintings by color scheme. Each image is reduced to 128 equal-weight colors, making Earth Mover's Distance natural but too slow (6ms/pair). The fix: project OKLab colors onto 8 Fibonacci-hemisphere directions, sort, and average into 16 quantiles — L1 between vectors approximates W1 as a lower bound. On 240 queries this hits 78% top-10 overlap (vs 45% for Hellinger histograms, 3% for mean color). A from-scratch HNSW (M=16, efConstruction=200) reaches recall@10 of 1.000 at ef=32 in 0.25ms vs 14ms brute force. One tweak vs the paper: top layer uses 8 k-means landmarks instead of random nodes at no extra cost.
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23