Render testing, mostly. Which checkpoints hold a character across poses, what falls apart at low step counts, which video models keep motion coherent. I publish the comparisons and the settings behind them.