Creating a Color Palette from an Image

Creating a visually pleasing and representative color palette from an image is a surprisingly complex challenge. Amanda Hinton details her iterative process to develop Spectrimage, a tool designed to extract five "human-selected" colors from any photograph. Her goal is to produce consistent, meaningful palettes without muddy averages or redundant swatches, balancing algorithmic precision with aesthetic intuition.

• Iteration 1: The RGB Struggle: Initially, Hinton used median-cut quantization in RGB, combined with complex partitioning and deduplication rules. While it "mostly worked," the system became a brittle "pile of patches" with thirteen named constants and six rules for gray, lacking reasonability and struggling with new edge cases.
• Iteration 2: OKLCH and K-means: Abandoning RGB's limitations (like HSL S blowing up near black), Hinton pivoted to the OKLCH color space, which offers a better measure of "colorful-ness" (chroma). She implemented K-means++ clustering (with K=10, then merging down) to group similar pixels, using a deterministic seed for consistency. Initial merging based on distance faced issues with distinct colors collapsing.
• Iteration 3: Hue-Weighting and Increased K: Recognizing that human perception weighs hue and chroma differently from lightness, Hinton introduced hue-weighting in the merge step. The chromatic plane now counts twice as much as lightness, preventing distinct hues from merging prematurely. She also increased K to 14, improving the detection of small, vibrant accents in largely achromatic images.
• Iteration 4: Refined Guards and Allocation: To address persistent failures (phantom colors, unwanted browns, redundant grays), Hinton added structural changes:
  • Phantom Guard: Clusters below 2.5% pixel weight and low chroma are dropped.
  • Mass Split: Achromatic and chromatic slots are allocated proportionally based on image content, merging same-bucket achromatic clusters to free up chromatic slots.
  • Centroid-Aware Representatives: Swatch selection now varies; for chromatic clusters, the highest-chroma pixel is chosen, while for essentially gray clusters, the pixel closest to the centroid ensures a true gray swatch.
• Future Improvements: Hinton notes further areas for refinement, such as thinning same-family ramps and improving sorting.

The result is an algorithm that produces human-perceived palettes more effectively, demonstrating how careful choice of color space, clustering, and nuanced merging rules can transform a difficult computer vision problem into an elegant solution.