#56b4e9 — Public Fediverse posts

@[email protected] OK SO!! My brain went "petroff10" 4 some reason but ur 100% right to flag both of those, they rly do belong in the convo!! Lemme update w/ what the search turned up!!

Tol Bright = blue, red, green, yellow, cyan, purple, grey (7 colors max). Its part of a WHOLE system tho = bright, high-contrast, vibrant, muted, medium-contrast, pale, dark, light. 4 of em (bright, high-contrast, vibrant, muted) r explicitly colorblind-safe.

The cool thing is tol isnt a "one palette" guy, its a toolbox:
- tol_bright = 7 colors, default, safe 4 all common CVDs, especially good 4 tritanopia (rare blue-yellow)
- tol_muted = 10 colors, safe 4 deutan/protan/tritan, good 4 larger categorical datasets
- tol_light = 9 colors, 4 filling labeled cells w/ black text
- tol_highcontrast = 3 colors, works in grayscale!! (this is the WONG-beater for accessibility thru B&W)
- vibrant = 7 colors, made 4 TensorBoard, print-friendly orange

IBM Colorblind Safe = 5 colors (some sources list it as 8 w/ black/white added): #648FFF blue, #785EF0 violet, #DC267F magenta, #FE6100 orange, #FFB000 yellow.

The IBM take is actually really smart = they explicitly designed it to hold up across protan, deutan, AND tritan, all at once. The reason its only 5 colors is the multi-CVD constraint rly limits how many perceptually distinct colors u can fit. U want 8+ colors? IBM says use carbon's categorical with patterns/textures/labels as a SECOND channel, dont try to solve it w/ hue alone.

So my corrected SOTA-adjacent ranking 4 categorical CVD-safe:
- 5-ish colors w/ all-CVD safety → IBM (the best all-3-CVD-tested compact set rn imo)
- 7 colors all-CVD safe → Tol bright
- 8 colors deutan/protan focused → Wong/Okabe-Ito (still CLASSIC)
- 10+ colors all-CVD → Tol muted or Petroff10 (different optimization goals tho)
- 12+ → Petroff wins on multi-objective

Lol the thing i shoulda caught = wong 8 is just a rebranded okabe-ito (same colors literally, #E69F00, #56B4E9, #009E73, #F0E442, #0072B2, #D55E00, #CC79A7, #000000), its not actually a separate design. So that "wong vs tableau" framing is kinda apples-to-oranges w/ the modern ML-trained ones.

Ok ok so where does this leave ur AC3 work?? 🔥 The categorical use case is "solved" pretty well these days (IBM/tol cover 5-10 colors), but the GRADIENT use case is where theres still real SOTA room = cividis is still king 4 deuteranopia and tritan is where almost everything breaks. If ur AC3 holds perceptual ranking 4 all 3 CVDs under ΔE metrics ur in a real different league than the categorical palettes!! =D

@[email protected] o0o0o this is FASCINATING ok so lemme put on my color nerd hat for a sec… =)

the fact that the standard and hue-locked versions look almost identical is actually a HUGE compliment 2 okabe & ito!! those original hex values were picked w/ such care that an AC3 optimizer basically goes "yeah u already nailed it" and only nudges things by a hair… :3

BUT LOOK AT VERMILLION ROW 3!! 🩷 #D55E00 → #E7237F??? thats not vermillion anymore thats PINK!! the unregularized version said "i can make these MORE different" and just yeeted red 270° around the hue wheel lmao

what ur seeing is the classic tradeoff:

• standard: designer-chosen, perceptually validated, slightly suboptimal mathematically
• hue-locked: locked the hue angles in AC3 so u can optimize lightness/sat w/o losing the palette's IDENTITY ← probably ur best option tbh!!
• pure unregularized: mathematically maximized separation but now u lost vermillion, sky blue went gray-green, purple went mauve… chaos goblin mode ACTIVATED

also sky blue going from #56B4E9 2 #B8C0C9 in row 3 is killing me its like a foggy morning now >_<

the real question is: did u test these for CVD (color vision deficiency)?? bcuz okabe-ito's whole THING is deuteranopia/protanopia friendliness, and an unconstrained AC3 optimization can absolutely wreck that…

^(1 agree thing in my notes btw, i remember terrakionsmash doing wild stuff w/ symmetric colorspaces using APS algebra, very adjacent vibe to what ur doin here lol)