T4GS: Learning Temporally Aware Compact Gaussians for Feed-Forward 4D Reconstruction

Checkpoints for T4GS (Temporally Aware Compact 4D Gaussian Splatting), a feed-forward 4D reconstruction model that decodes a compact set of timestamp-conditioned learnable query tokens into 3D Gaussians from a monocular video, without input camera poses.

Code: https://github.com/cvlab-kaist/C4G

Files

File	Description
t4gs.ckpt	T4GS model used in the paper (weights only, no optimizer state). Relative time embedding with sinusoid period 100, 2048 Gaussian queries, 224×224 input.
t4gs_vdm_refinement.safetensors	Video diffusion module (Wan2.1-VACE-1.3B based) for camera-controllable video generation.

t4gs.ckpt stores the encoder weights under state_dict with the encoder. prefix. Use it with the repository, e.g. to initialize training: bash scripts/train.sh model.encoder.pretrained_weights=/path/to/t4gs.ckpt.

The reconstruction model runs standalone: it predicts the Gaussians and renders novel views without the video diffusion module. The paper's novel-view-synthesis and tracking numbers are from the reconstruction model alone; the diffusion module is only used for the camera-controllable video generation application.

Lineage and license

t4gs.ckpt is initialized from the C3G static model (honggyuAn/C3G), whose backbone comes from VGGT (facebook/VGGT-1B, CC BY-NC 4.0), and was trained with supervision from MoGe-2 and CoWTracker. The weights are therefore released under CC BY-NC 4.0 (non-commercial use only). The code is MIT-licensed except for bundled third-party components; see LICENSE and THIRD_PARTY_LICENSES.md in the code repository.

Citation

@article{kim2026learning,
  title={Learning Global Motion with Compact Gaussians for Feed-Forward 4D Reconstruction},
  author={Kim, Mungyeom and Jeon, Minkyeong and An, Honggyu and Jung, Jaewoo and Ko, Hyuna and Han, Jisang and Yu, Hyeonseo and Shin, Donghwan and Hong, Sunghwan and Narihira, Takuya and others},
  journal={arXiv preprint arXiv:2605.31595},
  year={2026}
}