SenseNova open-sourced a 27B model that makes videos by writing code, and published full production logs too

Overview

It's called Code2Video. Rather than generating pixels the way video models do, it writes a program for each shot, renders it, looks at the frames and revises the code. I think the reasoning holds up for a certain kind of video. If a ball bounces, the bounce comes from the physics in the code, not from what bounces usually look like. Text, formulas and chart numbers get typed in from the source, so they don't come out as gibberish. And since every element is a named object, changing one t

SpinGraph analysis pending — check back after processing.