A deep dive into whether RGB integer values should be normalized by dividing by 255 or 256 when converting to floating point for image processing. The standard approach (÷255) maps 0→0.0 and 255→1.0, matching GPU conventions, but results in half-width bins at the extremes and inexact floating-point values. The alternative (÷256 with 0.5 bias) is a mid-tread quantizer that places values exactly at bin midpoints and offers marginally lower reconstruction error. The analysis frames both as uniform scalar quantizers (mid-riser vs mid-tread), concludes that for processing third-party images the standard ÷255 is correct, and that ÷256 only makes sense when you control both encoding and decoding. Mixing the two approaches is always wrong.
254 Impressions