Peking University Unveils UniMotion: A Unified Multimodal Framework for Human Motion
机器之心 · wechat · 2026-08-15
Researchers from Peking University and collaborators proposed UniMotion, integrating continuous human motion as an independent modality into a Unified Multimodal Model (UMM). The paper has been accepted by ECCV 2026.
- Architecture: Built on Show-o21.5B, using discrete tokens for text and continuous latent representations for Motion and RGB, coordinated via Hybrid Attention and modality-routed LoRA.
- Innovations: Introduces CMA-VAE for motion representation aligned with visual semantics (removable at inference) and LRA (Latent Reconstruction Alignment) for self-supervised pre-calibration.
- Capabilities: Unifies seven tasks including motion understanding, generation, prediction, editing, and visual human pose recovery.
- Performance: Shows competitive results on datasets like HumanML3D, demonstrating better semantic execution in complex motion generation compared to specialized baselines.
More from Multimodal
- How meloform works: WASM pitch detection plus differentiable circuits learning 1,022 sine coefficients — henloitsjoyce · 2026-10-16
- Documentary filmmaker urges peers to watch 'You Can See Everything' — kyliebytes · 2026-10-16
- Redditor's first Cumfy/Seedance video nails a cinematic one-shot generation — Wave_N_Wind · 2026-10-16
- Kling 4 returns after 8 months: 10 hyper-realistic AI ads for under $5 — CurieuxExplorer · 2026-10-16
- Shanghai AI Lab unveils Intern Lumina U2, a diffusion LLM unifying image, video and 3D tasks — Shanghai-AI-Laboratory · 2026-10-16
- NAVER AI Lab proposes Speculative Forcing to speed up autoregressive video generation — naver-ai · 2026-10-16