Should you normalize RGB values by 255 or 256?

4 hours ago

The article compares two methods for converting 8-bit integer image data to floating-point: standard division by 255 and alternative division by 256 with a 0.5 bias.
The standard approach maps integer 0 to 0.0 and 255 to 1.0, ensuring black is 0.0, but it results in stretched bins at extremes, making extreme values less frequent in output.
The alternative approach places floating-point values midway between integers, offering exact midpoint positions and potentially better dithering, but ties logic to 8-bit dynamic range.
From a quantization perspective, the standard method is a mid-riser quantizer with L=255, while the alternative is a mid-tread quantizer with L=256, affecting reconstruction error and bandwidth usage.
For processing images from unknown sources, division by 255 is recommended for compatibility; division by 256 may be considered only if controlling both saving and loading, despite minimal precision gains.

Hasty Briefsbeta