WorldFlow is a paper on a generative world model that predicts future video and 3D object trajectories together.
Method: The 3D motion is parameterized as continuous displacement tokens aligned with the video grid, and one Mixture-of-Transformers backbone denoises both modalities under a rectified-flow objective.
Training: The model is trained on 1.44 million object-level samples spanning simulated and real-world domains.
Results: On the SpatialCog set, the paper reports an ADE of 0.060 m, a 54.2% reduction against the strongest external baseline. Compared with its flow-only variant, ADE and FDE fall by 55.9% and 68.5%.
Benchmark: The authors introduce WorldFlowBench, with 1,678 evaluation samples from 11 data sources.
Sign in to comment.
TL;DR of the paper, generated automatically.
WorldFlow is a paper on a generative world model that predicts future video and 3D object trajectories together.