KEYPOINT-GUIDED 4D GAUSSIAN SPLATTING WITH DECOUPLED SPATIO-TEMPORAL FLOW REFINEMENT

Jusheng Zhang, Jinzhou Tang, Zhuojie Yang, Sidi Liu, Kaiyu Wu, Mingyan Li, Jian Wang, Keze Wang, Yufeng Yang

Keywords: Keypoint, 4D Gaussian Splatting

TL;DR: This study proposes a motion deformation method guided by key point radiation features, which achieves efficient and high-quality dynamic 4D Gaussian point cloud generation.

Abstract:

We propose KG4D, a novel method for generating time-aware 4D representations
from a single static image or video. Previous methods largely rely on weak su-
pervision signals, failing to introduce fine-grained supervision necessary for cap-
turing detailed spatio-temporal dynamics. In contrast, our approach employs Har-
monic Spatio-temporal Encoding (HSE) to achieve efficient spatio-temporal sep-
aration during training, allowing the model to represent dynamic scene changes
more accurately. Furthermore, Keypoint Feature Calibration (KFC) ensures pre-
cise pose consistency, and Wasserstein Gradient Flow (WGF) enhances motion
coherence, effectively reducing artifacts. Comprehensive evaluation and ablations
demonstrate that our proposed KG4D outperforms existing state-of-the-art meth-
ods on various benchmarks in dynamic 4D generation and novel viewpoint syn-
thesis, validating its effectiveness and superior generation capability.

Primary Area: generative models

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: 9002