ABOUT
Everyone in the room shares one jukebox queue and the same three televisions, which stay in sync so a late arrival lands at the right moment of the tape. Voice chat is spatial, so people get louder as you walk toward them. Visitors can upload photos, GIFs, and short videos, pick them up, carry them across the room, and hang them wherever they like.
BUILD
The room is built with React Three Fiber. Presence, chat, and voice signalling run through a PartyKit room that sends one merged snapshot ten times a second, so thirty people cost the same to broadcast as two. Voice is a peer-to-peer WebRTC mesh, with a relay for people on strict networks.
PERFORMANCE
The first model weighed 122 MB. Re-encoding its textures and geometry brought the download to about 13 MB, and baking ambient occlusion in Blender replaced a pass that used to run every frame. Dropping real-time glass refraction and limiting the floor reflection took the time to draw a frame inside the diner from 41.7 to 8.3 milliseconds on an M3 MacBook Pro.
CREDITS
Almost nothing in the room was modelled by me. The furniture, jukebox, and fixtures are 29 Creative Commons models by Sketchfab artists, each credited by name on a plaque inside the diner.