PixVerse launches R2 world model that learns from failures, cutting long-horizon drift by 35.8%

LearnWithBishal · x · 2026-09-26

PixVerse has released its R2 world model with a technical report, positioning it as a step toward neural game engines. Its key feature is the Error Bank: the model stores failure states, replays them during training, and learns to recover mid-generation. Internal tests show a 35.8% reduction in long-horizon drift, with quality improving the longer you play.

Original post →

More from Multimodal

Multimodal channel →