doing the one offs is how you learn how to build the general system

my intuition is yes we do lose something, but it doesn't matter unless the image in question is very low resolution. every texel subtends a non-infinitesimal patch on your integration domain, but when the image is huge you can handwave that away.

Oh man what were you even working on

defining convolution through spherical harmonics feels... circular? I'm also not sure if it's possible because the nice formulas exist for symmetric kernels but I don't know about the general case

I'm just saying that even finding the simple formal definition of what it _means_ to convolve two spherical functions seems weirdly difficult

yeah even riemann sums on the sphere are weird because how do you compute the area of sphere surface patches? i almost had to write a whole section about it in my blog post but luckily there was already a really good explanation www.rorydriscoll.com/2012/01/15/c...

is spherical convolution (as in, convolving functions on a sphere) really such a niche operation? the only definitions i can find are all 10 pages deep in some ml papers OR on someone's lecture slides from 2015. you'd think the wiki article on convolution would at least mention something...

the internet has ruined us
i mourn my innocence

probably, inverted normals😅

still cooking... another one of those posts that got out of control...

doing the yearly hanami at the uw quad

I think "male loneliness crisis" is really referring to "male singleness crisis" :)
which could potentially be cured by hanging out with other dudes (the larger your friend group, the more chances there are to meet a partner), buut it might not be the most efficient way

ngmi. you should be running a dark software factories with 10000 parallel agents handling tickets for your agentic SaaS.

X Japan - Rusty Nail (Anime Version) YouTube video by Yoshiki

type of shit that you see on the tv or some vhs tape once, when you're 9 years old, and then spend your teenage and young adult years wondering whether it was some fever dream or implanted memory, until one day you stumble upon it on youtube
www.youtube.com/watch?v=M1dc...

the point here is that uefn doesn't need a programming language

girugamesh 🤘

your renders are always so slick

tbh, it's probably a bit inflated because of some of the math begin split out into separate paragraphs for better readability. but by pure pagecount it's shaping up to be the record holder (current king is one of the assembly tutorials clocking in at 28 pages)

...and im still not done yet...
(no ai was used)

something something american healthcare
something something social safety net

the laughs were worth the few braincells lost 😁

omg link please

yeah, this is how i found out i was using the wrong coefficients (there was more than 1 wrong) :D

this stuff is a minefield, i've found at least 2 other typos in another source - getting spherical-to-cartesian conventions wrong, and the coefficients incorrect. makes you wonder whether the replication crisis in science is really a typo crisis.

pretty sure Green's SH lighting paper (or at least the version i have) has a typo: l=2 m=2 should have a factor of 1/4 not 1/2. it's 3 vs 1: Sloan's paper, wikipedia and my own derivation agree.

alright let's add one more band

good suggestion, it didn't even occur to me and so many visualizations already use red/green!

oh yeah its all comin together

an oddly satisfying illustration why the number of spherical harmonic functions up to and including a certain degree is always a square.