OmniDirector: General Multi-Shot Camera Cloning without Cross-Paired Data

A new framework called OmniDirector enables video generation models to clone camera motion from reference footage without requiring cross-paired training data, according to a paper posted to arXiv on June 11, 2026 [1][2]. The work addresses a persistent challenge in AI-driven video synthesis: existing camera-motion cloning techniques either rely on parametric representations that break down during multi-shot sequences or require synthetic cross-paired datasets that are scarce and limit performance in complex scenes [2]. OmniDirector introduces a general camera motion representation that encodes camera parameters as grid motion videos, a visual format that can integrate diverse trajectories for multi-shot generation [1][2]. The framework was trained on roughly one million camera grid-video pairs, a scale the authors describe as necessary to coordinate characters, actions, and cameras within multimodal diffusion transformers [1][2]. A hierarchical prompt expansion agent was designed to harmonize different control signals by systematically describing camera motion and visual content through an understanding of signal relationships [2]. The paper states that extensive experiments demonstrate superior performance and controllability compared to prior methods [2]. The preprint appeared on arXiv, the open-access repository that hosts electronic preprints across physics, computer science, and related fields [6]. As of late 2024, arXiv was receiving approximately 24,000 new articles per month and had surpassed two million total submissions [6]. The OmniDirector paper is accompanied by a project page and is listed with standard arXiv Labs integrations, including bibliographic tools and code-finding services, which are community-developed features that operate under arXiv’s framework for third-party collaborations [4][5]. arXiv Labs, launched in 2020, provides a formalized space for individuals and organizations to build experimental tools on top of the repository’s article pages [5]. The initiative requires partners to adhere to arXiv’s stated values of openness, community, excellence, and user data privacy, and collaborators receive only minimal, anonymized user data necessary for feature functionality [5]. Current Labs offerings include citation-tree explorers, recommender systems, and links to code repositories [4]. The OmniDirector listing surfaces several of these integrations, though the paper itself has not undergone peer review, consistent with arXiv’s moderation-but-not-peer-review policy [6].