Microsoft Research’s World-R1 Uses Flow-GRPO and 3D-Aware Rewards to Inject Geometric Consistency Into Wan 2.1 Without Architectural Changes

Microsoft Research and Zhejiang University introduced World-R1, a post-training RL framework for Wan 2.1 that improves 3D consistency in text-to-video generation.

World-R1 uses Flow-GRPO and rewards from 3D reconstruction, camera trajectory matching, and visual quality scoring to encourage coherent geometry across frames.

The approach keeps the base model and inference cost unchanged, avoiding expensive architectural changes or explicit 3D supervision.

It targets a key weakness in video models—scene instability under camera motion—making generated videos more physically consistent and practical.