Orator: LLM-Guided Multi-Shot Speech Video Generation
Keywords: Speech video generation, Multimodal video generation, Human video dataset, LLM-directed human video synthesis
Abstract: In this work, we propose a novel system for automatically generating multi-shot speech videos with natural camera transitions, using input text lines and reference images from various camera angles. Existing human video generation datasets and methods are largely centered on faces or half-body single-shot videos, thus lack the capacity to produce multi-shot full-body dynamic movements from different camera angles. Recognizing the lack of suitable datasets, we first introduce TalkCuts, a large-scale dataset containing over 500 hours of human speech videos with diverse camera shots, rich 3D SMPL-X motion annotations, and camera trajectories, covering a wide range of identities. Based on this dataset, we further propose an LLM-guided multi-modal generation framework, named Orator, where the LLM serves as a multi-role director, generating detailed instructions for camera transitions, speaker gestures, and vocal delivery. This enables the system to generate coherent long-form videos through a multi-modal video generation module. Extensive experiments show that our framework successfully generates coherent and engaging multi-shot speech videos. Both the dataset and the model will be made publicly available.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 608
Loading