UniAction

UniAction is a unified action-conditioning framework for video-based world models, described in a 2026 arXiv preprint by Bowen Xue, Zizhang Li, Hong-Xing Yu and Jiajun Wu. It represents any action as a sparse set of action elements, each with a discrete semantic identity and a continuous spatiotemporal grounding coordinate. These coordinates are mapped into the pretrained video model's native positional space, so different action types share the same spatiotemporal structure.

The method adds cardinality-calibrated attention to handle bias from differing action representation granularities. It is evaluated on four action forms: force, camera control, hand pose and robot pose. According to the authors, it follows actions more faithfully than visual or token-based conditioning, and a single jointly trained model outperforms action-specific expert models. The project page also reports that multiple action controls can be combined at inference even though only one action type is seen at a time during training, and that it generalizes to new scenes and timings.

A YouTube creator reports that the model is built on Wan 2.2 and that code, training and evaluation scripts have been released. The provided pages do not give a repository link, weights location or license, so those are left unspecified. It is primarily a research release rather than a hosted product.