Avatar generator with color you can defend

Yes, if it repeats the same three steps: normalize to NFC, trim and collapse whitespace, lowercase, then FNV-1a 32 over the UTF-8 bytes. The byte encoding is where implementations usually part company — hashing UTF-16 code units instead of UTF-8 gives a different number for any name outside ASCII, which is why one account can end up green in the browser and blue in a generated email. Each avatar here shows its hash as 0x-prefixed hex, so you can print the same value from Python, Go or Postgres and check the two agree before you ship.

That is arithmetic rather than a bug: twelve plates over five names collide 62% of the time, and over ten names it is a near certainty. Color is a recognition aid — it makes your own row findable in a list you scroll every day — and the initials are what actually distinguish two accounts. If you need the picture itself to differ, switch to the identicon: 15 independent bits give 32,768 patterns, or 393,216 combinations once the plate is counted.

Because that is the only choice that cannot fail. Black or white on any sRGB color never drops below 4.58:1 — the worst case is the crossover at L* 49.4, where black and white score the same — so picking whichever is better with bestTextColor guarantees WCAG AA for the initials on every plate the palette can produce. A tinted ink taken from the plate's own hue looks more designed and drops under 4.5:1 on the light half of the wheel, which is how avatar initials end up illegible on exactly the pale mint tile someone screenshots for the bug report.

For the three dichromacies, yes, and the page proves it rather than claiming it: the palette is built by a greedy walk that rejects any candidate plate landing within ΔE76 12 of an accepted one after protanopia, deuteranopia and tritanopia have been simulated, and the panel under the grid reports the closest surviving pair for each. Achromatopsia is the honest exception. Once hue carries nothing, only lightness is left, and at the same spacing exactly 2 of the 12 plates survive — no palette of a dozen colors can be safe there, which is the argument for the identicon pattern as a second channel.

That is a renderer ignoring dominant-baseline, and the file here avoids it. librsvg — what ImageMagick, GTK apps and several server-side thumbnailers use — does not implement dominant-baseline at all, so an avatar centered with dominant-baseline="central" comes out with the letters resting on the vertical midline instead of straddling it. These files position the text with dy=".35em" from a y of 50 instead, which every SVG renderer honors because it is plain text positioning rather than a baseline hint.

Four rules cover the cases that break a naive first-letter-of-each-word split. Sixteen nobiliary particles — van, von, de, della, ibn and the rest — are dropped when two other tokens remain, so Ludwig van Beethoven is LB and not LV. A name whose first character is Han, Kana or Hangul returns that one character, because a two-character initial reads as a different word in those scripts. Letters are cut by grapheme with Intl.Segmenter, so an é built from a combining mark or a flag emoji stays whole instead of splitting into halves. And a line with no letter, digit or emoji at all falls back to the identicon and says so, rather than drawing an empty tile.

The id, in almost every case. A display name is editable, and the day someone adds a middle initial their avatar changes color and their colleagues stop finding them in a list — the same failure as rehashing on every sign-in. Hash the immutable primary key and pass the display name separately for the initials; this page keeps them separate for that reason, showing the hashed key beside the letters it drew. Locale is the other trap: use toLowerCase and not toLocaleLowerCase, because the locale-aware form maps the Turkish dotted capital İ differently and a Turkish browser would then compute a different hash than your server.