Notes on color science, pixel formats, and LLM inference.
Technical notes from Polyverse's graphics and inference work: original research, reference values, and where standards and common practice quietly diverge.
Research our own work and measurements Reference standards and derivations, worked through Review other people's work, explained and compared
Blue Noise Token Sampling
Sampling error is quantization error. How dithering research became a generative blue noise RNG, and what happens when llama.cpp token sampling stops drawing white noise.
Read article →
From Stateful to Stateless Blue Noise
Five states, seven, one bit, none. Four forms of the blue noise tree generator, and the two chance findings that led to them. The smallest stores nothing and can compute any sample directly.
Read the draft →
Chroma Rotation Averaging
Match the colors, not the channels. Histogram matching along six directions of the chroma plane, so an upscaler's drifted colors come back without its dark blues turning purple.
Read article →
Integer Color Format Conversion Relations
Bit replication, truncation, and float conversion between integer color formats. Why uint8 and uint16 are canonical, and why rounding during truncation is incorrect.
Read article →
Color Space Specification
Authoritative definitions and derived values for sRGB, CIELAB, OKLab, Y′CbCr, and related color spaces. What each standard actually defines, and what is merely derived.
Read article →
The D65 Whitepoint
Precision, history, and derivation of the standard daylight illuminant. Temperature scales, the daylight locus, and which chromaticity values are authoritative for which purpose.
Read article →
Get the numbers right
From dithering to LLMs
Articles were written using AI assistance and fully verified against the relevant standards and sources.