Microsoft Announces World-R1

World-R1 injects 3D constraints via reinforcement learning without changing the architecture, in order to reduce geometric inconsistency in video generation models.

To achieve this, it builds a pure text dataset for world simulation, and in Flow-GRPO uses feedback from a pretrained 3D foundation model and a vision-language model as rewards.

The reward combines a 3D-aware reward and an aesthetic reward to align structural consistency and visual quality at the same time. On top of this, periodic decoupled training is added to alternately tune geometric alignment for static scenes and motion flexibility for dynamic scenes.

In evaluation, it is presented as improving 3D consistency while maintaining the visual quality of the existing foundation model.