There is no temporal accumulation thanks to the limitations of VRChat, but if you can leverage that then I highly recommend it! Just 4 frames is typically enough, check the rockstar video on Red Dead Redemption 2 for more information

I sourced a few resources:
media.contentapi.ea.com/content/dam/...
advances.realtimerendering.com/s2015/The%20...
youtu.be/9-HTvoBi0Iw&...

Also, the steps are dithered using a 2x2 bayer matrix and then averaged over a 2x2 kernel blur so the result is effectively half res

In all, most parts of the screen spend less than 10 steps

It is marched, the clouds are quite solid which allows the march to stop early once it becomes fully opaque, and there is a binary search before the detailed march to find the cloud shape so empty regions waste few steps too

Also thanks to @jadenseraph.bsky.social for all of the sources and feedback while I make this. It'd look doo doo without them

Here are some ugly early progress screenshots from the first 2 hours

Find on Gumroad: fursonify.gumroad.com/l/spirivalip...

Find on Jinxxi: jinxxy.com/KufiBean/spi...