Pixels are read as sRGB, linearised with the piecewise IEC 61966-2-1 transfer function, converted to CIE XYZ under a D65 white point, then to CIELAB. Averaging happens in linear light, which is why a 25×25 aperture over a noisy shadow gives a usable answer instead of a value biased bright by gamma. A single pixel is a measurement of sensor noise as much as of subject color. For identifying a garment, a paint chip, or a skin tone, 5×5 or 11×11 is the honest reading. A browser canvas is 8-bit, so a JPEG loaded here gives roughly 13 distinct code values across the darkest stop. With the sidecar running, raw and 16-bit TIFF arrive as a 16-bit buffer and the same stop carries several thousand. Shadow colour identification is the measurement that changes most. On the browser path the pixels arrive already converted to the display path. A file tagged Adobe RGB or Display P3 reads correctly in current Chrome and Safari; an untagged wide-gamut export reads as sRGB and the saturated values will be wrong. Tag your exports. On the sidecar path this cannot happen — the working space is stated explicitly and its primaries are used for every conversion.

Nothing. There is no account, no upload, no analytics, and no advertising. The page makes no requests to any other origin — the typefaces are served from here or fall back to the ones already on your machine, so no font CDN sees your visit either. After the first visit the tool is cached and runs with the network off. You can install it from the browser menu and it behaves like a normal application. No paid tier, no feature held back. If it saved you an afternoon and you want to send a coffee, there is a link in the sidebar; if not, that is completely fine and nothing changes. If a measurement looks wrong, or a raw file will not open, say so — that is the most useful thing anyone can send. Contact. Measurements are only as good as the file you feed in. For work where the numbers carry consequences, verify against a calibrated target rather than trusting any single tool.

python dyelot_server.py starts a loopback-only decoder built on LibRaw. Files are read from memory, never written to disk, and never leave the machine. LibRaw would otherwise rescale each frame to a pleasing exposure, which silently destroys any comparison between two files. Every decode here is at the raw's own scaling. When the output is being reduced by 2× or more anyway, each 2×2 CFA block is read directly instead of being interpolated and then averaged away. Fewer invented colours near edges. Blown and crushed pixels are counted at full decode resolution, before any downsampling can average them into something that looks recoverable. As shot, daylight, auto, and none. "None" applies unity multipliers and shows the sensor's own balance, which is where a colour cast that survives every edit usually turns out to live.

ΔE00 — CIEDE2000 difference. Under 1.0 is invisible to most viewers, 1–2 is visible on a side-by-side edge, above 5 reads as a different color. Human skin across every complexion clusters near 123° in the Cb/Cr plane; melanin changes the distance from centre, not the angle. A face sitting off that line is a white-balance problem, not a skin problem. The CCT shown for a sample is McCamy's approximation with Ohno's Duv, applied to the pixel as rendered. It answers "what illuminant would produce this patch if the patch were neutral" — useful when you click a grey card, meaningless when you click a red jacket. Each tonal zone is averaged as a vector in a*/b*, so a frame with green shadows on one side and magenta on the other reports near zero rather than a false cast. Mean absolute chroma is reported alongside to catch that case.