Evaluation on our synthetic TrueBones dataset [4].
We compare our method against SOTA dynamic monocular reconstruction methods: (i) optimization-based methods HiMoR [1] and OriGS [2], (ii) generative-based Cog-NVS [3]. The top video depicts the pre-scan and dynamic sequences, on which all models are trained on. Each row contains renders of all methods on a specific test camera alongside with the ground-truth.
Our method drastically outperforms the baselines in maintaining a consistent object geometry and sharp appearance, while accurately modeling the scene dynamics.
Training Video
Show Pre-scan
Results
[1] Yiming Liang, Tianhan Xu, Yuta Kikuchi. HiMoR: Monocular Deformable Gaussian Reconstruction with Hierarchical Motion Representation, CVPR 2025
[2] Junyi Wu, Jiachen Tao, Haoxuan Wang, Gaowen Liu, Ramana Rao Kompella, Yan Yan. Orientation-anchored Hyper-Gaussian for 4D Reconstruction from Casual Videos, NeurIPS 2025
[3] Kaihua Chen*, Tarasha Khurana*, Deva Ramanan. Reconstruct, Inpaint, Test-Time Finetune: Dynamic Novel-view Synthesis from Monocular Videos, NeurIPS 2025