AI Navigate
インサイト最新記事AI大全カオスマップ

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

MarkTechPost / 5/1/2026

📰 NewsIdeas & Deep AnalysisModels & Research
Read original →
共有:

Key Points

  • Microsoft Research’s World-R1 uses reinforcement learning to improve geometric (3D) consistency in text-to-video outputs.
  • The approach employs Flow-GRPO and 3D-aware reward signals to encourage stable 3D structure across generated frames.
  • It reportedly enhances 3D consistency in Wan 2.1 without requiring any architectural changes to the underlying text-to-video model.
  • The work targets a common weakness in generative video models: inconsistency that can break geometric plausibility over time.

Microsoft Research's World-R1 Uses Reinforcement Learning to Force 3D Consistency Into Text-to-Video Models

The post Microsoft Research’s World-R1 Uses Flow-GRPO and 3D-Aware Rewards to Inject Geometric Consistency Into Wan 2.1 Without Architectural Changes appeared first on MarkTechPost.

関連おすすめサービス

※当サイトはアフィリエイト広告を利用しています
🎧

Notta搭載AI議事録イヤホン ZENCHORD1

AI時代の仕事術。Notta搭載で会議の議事録を自動生成するスマートイヤホン。

🎙️

AI搭載ボイスレコーダー Plaud

世界100万人が愛用。AIで文字起こし・要約を自動化するボイスレコーダー。

🖼️

画像高画質化AIツール Aiarty Image Enhancer

AIで画像を高画質化。写真・イラストを簡単にアップスケール。