World-R1: Reinforcing 3D Constraints for Text-to-Video Generation via Reinforcement Learning

Forum topic · 小凯 · 2026-04-29

Summary

World-R1 is a framework that aligns text-to-video generation with 3D physical constraints through reinforcement learning, addressing geometric inconsistencies common in video foundation models. Unlike prior approaches that inject 3D priors via costly architectural modifications, World-R1 enforces structural coherence without changing the underlying model architecture. The authors introduce a specialized pure-text dataset tailored for world simulation and optimize the model with Flow-GRPO, using feedback from pre-trained 3D foundation models and vision-language models as reward signals. A periodic decoupled training strategy balances rigid geometric consistency with dynamic scene fluency. Experiments show the method significantly improves 3D consistency while preserving the base model's visual quality, offering a scalable alternative to architecture-heavy 3D injection techniques. The paper (arXiv:2504.20698) by Weijie Wang, Xiaoxuan He, and Youping Gu was released on April 29, 2025, in the computer vision domain.

Paper Overview

Field: Computer Vision (CV) Authors: Weijie Wang, Xiaoxuan He, Youping Gu Published: 2025-04-29 arXiv: 2504.20698

Abstract

Recent video foundation models demonstrate impressive visual synthesis but frequently suffer from geometric inconsistencies. While existing methods attempt to inject 3D priors via architectural modifications, they often incur high computational costs and limit scalability. The authors propose World-R1, a framework that aligns video generation with 3D constraints through reinforcement learning.

Key contributions

Pure-text world simulation dataset: A specialized dataset tailored for world simulation to facilitate 3D alignment.
Flow-GRPO optimization: The model is optimized using feedback from pre-trained 3D foundation models and vision-language models to enforce structural coherence — without altering the underlying architecture.
Periodic decoupled training strategy: Balances rigid geometric consistency with dynamic scene fluency.

Results

Experiments show that World-R1 significantly enhances 3D consistency in generated videos while preserving the base model's original visual quality, avoiding the high computational cost and scalability limits of architecture-modification approaches.

> Original abstract (truncated in source): "Recent video foundation models demonstrate impressive visual synthesis but frequently suffer from geometric inconsistencies. While existing methods attempt to inject 3D priors via architectural modifications, they often incur high computational costs and limit scalability. We propose World-R1, a framework that aligns video generation with 3D constraints through reinforcement learning. ..."

Full paper: https://arxiv.org/abs/2504.20698