MoRAE: A New Flow-Friendly Self-Supervised Latent Space for Text-to-Motion Generation
A new research paper, MoRAE, proposes a method to improve text-to-motion generation by creating a flow-friendly latent space from self-supervised representations. The paper, available on arXiv (2607.29180), addresses the failure of directly using Motion-JEPA as a frozen encoder for generative models. The authors diagnose two geometric bottlenecks: the JEPA feature space is spectrally ill-conditioned, causing unstable Gaussian-to-data transport, and flow residuals align with decoder-sensitive directions, amplifying small latent errors. MoRAE aims to overcome these issues, enabling more semantically correct, temporally coherent, and physically plausible motion generation. The work is relevant to the fields of AI and computer vision, particularly in animation and robotics.
Key facts
- Paper titled 'MoRAE: Flow-Friendly Self-Supervised Latents for Text-to-Motion Generation' is available on arXiv.
- The arXiv ID is 2607.29180.
- The paper addresses text-to-motion generation, focusing on semantic correctness, temporal coherence, and physical plausibility.
- It proposes using Representation Autoencoders (RAEs) with a frozen self-supervised encoder.
- Direct transfer using Motion-JEPA as the frozen encoder fails.
- Two bottlenecks identified: spectral ill-conditioning of JEPA feature space and alignment of flow residuals with decoder-sensitive directions.
- MoRAE aims to create a flow-friendly latent space to improve generation.
- The paper is categorized as a cross announcement on arXiv.
Entities
Institutions
- arXiv