How scoring works

Every judge and compare result includes these comparison signals.

• pixel_match is mean per-channel agreement from 0 to 100 after compositing both images over white. In foreground scope, RGB agreement is multiplied by alpha agreement.
• ssim is an in-house mean Structural Similarity Index Measure over a Gaussian window. In foreground scope, luma SSIM is multiplied by alpha-plane SSIM.
• combined is the equally weighted mean of pixel_match and ssim.
• scope is canvas by default. --foreground evaluates only the union of pixels that are non-transparent in either image and incorporates alpha agreement.
• regions are ranked bounding boxes produced by tiling the frame, measuring per-block error, and merging contiguous high-error blocks. Coordinates are in reference-image pixels.

Foreground scope is usually the right choice for transparent icons because it does not reward shared empty canvas. Canvas scope remains useful when the full background is intentionally part of both images.

High pixel match with lower SSIM often points to structural or edge misalignment; inspect a blend or side-by-side image. Lower pixel match with higher SSIM often points to palette differences; inspect the reference colors.

Scores do not measure semantic naming, editability, component decomposition, or whether a visually important local defect remains. Use inspect, visual artifacts, and human or vision-model review alongside the numbers.