Skip to content
Book Open access

SegMoDiff: Semantic-Faithful Text-to-Motion Generation via Segmented Raw-Sequence Diffusion

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 44 references

Abstract

Diffusion models have shown remarkable success in generating 3D human motion from text descriptions, demonstrating superior realism and diversity. However, directly modeling the full raw motion sequence is computationally intractable. Existing diffusion-based methods either compromise essential computations on the raw sequence or perform low-dimensional modeling in a latent space, inevitably introducing semantic deviations. To overcome this challenge, we propose SegMoDiff, a text-to-motion diffusion framework that enables adequate and fine-grained semantic modeling directly on raw sequences by partitioning them into tractable segments. SegMoDiff pivots on a segmented motion diffusion transformer (SegMoDiT), where overlapping contextual attention is introduced to preserve cross-segment coherence. On top of this, segment-wise text enhancement based on large language models is leveraged to achieve precise text-motion alignment. To further accelerate inference speed, we incorporate distribution matching distillation to reduce sampling steps. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art methods in semantic fidelity while delivering superior inference speed. The code is available at https://github.com/wjxxyz/SegMoDiff.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.