Qualitative results on the synthetic Spring-Gaus Dataset

All videos are rendered from novel views and present the full 30-frame sequence. Each model was trained on the first 20 frames of each 30-frame sequence with only three views.

Sparse-view results on the PAC-NeRF Dataset

All videos are rendered from novel views and present the full 14-frame sequence. Since this dataset is designed for physical parameter estimation, each model was trained on all 14 frames.

Qualitative results on the real-world Spring-Gaus Dataset