Opus 5.5 用函数逐帧生成视频不漂移

rompasso @0xrompasso原推发布于 2026.10.10

在 X 上打开原推

Pixel video models get worse the longer they run. Opus 5.5 does the opposite - frame four hundred costs exactly what frame four did. Nobody trained that into it with more video. Nobody bolted on a bigger effects dataset either. Instead of predicting the next frame from the last one, the model writes a single function that paints a given frame directly, no memory of what came before required. What that one change actually buys: → Every frame comes from the same function evaluated at a different moment, so there’s nothing to slowly go wrong → Nothing drifts, because there’s no previous frame being fed back in and degrading a little each time →A five-minute piece costs the same attention per second as a fifteen-second reel, so length stops being the risk → Fixing one typo or one color means editing the function once, not re-rendering a sequence that happened to come out right →The same math reproduces identical output on a different machine, which pixel-by-pixel video generation never guarantees Within hours of launch the timeline had full documentaries running five minutes long, not just showreels. Length stopped being the enemy. It became the proof. The interesting question was never why one clip looks good. It’s why the fiftieth frame doesn’t quietly get worse. How writing one frame as a function beats predicting the next one is in the article below 👇

这条是引用着另一条写的：x.com/0xcodio/status/2108134587633221885