KlingTeam Proposes Cross-Category Video Motion Transfer Framework with New Dataset
KlingTeam · hf · 2026-08-04
Existing video motion transfer methods rely heavily on fixed structural correspondence between reference and target objects, which fails when morphological differences are vast. KlingTeam introduces Motion Beyond Morphology, a perspective that transfers motion beyond fixed structures by preserving dynamics meaningful across diverse morphologies.
The proposed two-stage framework works as follows:
- Stage I: Learns multi-granularity abstract motion representations to bootstrap cross-category video pairs, providing supervision that preserves transferable dynamics.
- Stage II: Internalizes this supervision into direct reference-video-conditioned generation, eliminating the need for explicit motion extraction during inference.
Additionally, the team releases OpenVMT-Dataset and OpenVMT-Bench for training and evaluating image- and text-conditioned motion transfer across Same, Near, and Far category gaps. Extensive experiments demonstrate state-of-the-art motion fidelity and target preservation.
More from Multimodal
- Suno Music Meets Hailuo H3: Impressive AI Video Generation Workflow — techhalla · 2026-08-04
- AI-Generated Short Films for Imaginary Fashion Brand Troy — gizakdag · 2026-08-04
- Minimax H3 Local Test: 2MP Video Generation in 51 Minutes on RTX 6000 Pro — JahJedi · 2026-08-04
- MiniMax H3 launches on TapNow with native 2K output, up to 60% off until Aug 4 — SimplyAnnisa · 2026-08-04
- Minimax H3 Video Model New Demo Clip Released — JahJedi · 2026-08-04
- Open-Source LoRA for Satellite Image Editing Uses AI to Auto-Generate Training Data — LimitlessSaint · 2026-08-04