4DGS-WAM, an object-centric model for representing changing scenes, posted the highest reported scores in the paper's test of reconstructing past scenes and led the evaluated video baselines when it rendered future images from a supplied camera. But when the future ego camera had to be predicted as well, the paper says the model trailed Envision4D. The preprint evaluated the system on KITTI-MOT for future-observation prediction and reconstruction of observed frames.
A scene built from persistent pieces
4DGS-WAM maintains a decomposed, renderable 4D Gaussian state. In practical terms, it separates a persistent static bank from dynamic Gaussian splats tied to individual objects, so the scene can be rendered while those object-linked pieces change over time. For future rendering, the model aggregates the static background from past multi-view observations and renders it at the target camera pose.
The predictive part has two stages. A policy predicts target-horizon actor actions from object-centric trajectories. An action-conditioned world model predicts transformations of the associated Gaussian splats. The policy and world model are trained separately.
The perception stage uses vision foundation models for segmentation, optical flow, depth and camera poses. Target-time masks, object centers and correspondences are available during training, but not at inference. Thus, the system's training inputs include target-time perception signals that are withheld when it makes a prediction.
Two ways to judge a scene model
The benchmark separates future prediction from past reconstruction. Future prediction is summarized as mean±std over three sequence-level means at horizons h=1 and h=3, corresponding to six frames per method; all methods use the same prefixes and target frames. Past reconstruction is scored on observed frames at given past cameras and summarized separately as mean±std over three sequences.
The future-image scores use three image-comparison measures: PSNR and SSIM, where higher values are favorable in the reported comparison, and LPIPS, where lower values are favorable. With the given future camera, the policy version of 4DGS-WAM scored 18.80±1.21 PSNR, 0.596±0.091 SSIM and 0.161±0.042 LPIPS. The paper states that it led the video baselines on all three measures.
When the ego camera was predicted rather than supplied, the reported result was 16.47±3.05 PSNR, 0.482±0.033 SSIM and 0.245±0.107 LPIPS. The paper's summary says 4DGS-WAM trailed Envision4D in that comparison. The two results are different tests: one gives the model the target viewpoint, while the other uses a predicted vehicle camera.
The reconstruction lead needs context
For the past-reconstruction task, 4DGS-WAM reported 27.63±2.07 full-frame PSNR, 0.888±0.030 SSIM and 0.053±0.012 LPIPS. On dynamic regions, it reported 25.53±3.58 PSNR and 0.025±0.006 LPIPS. The paper describes these as the highest reported full-frame and dynamic-region metrics, averaged over three sequences.
That ranking comes with a rendering caveat. In 4DGS-WAM's quantitative renders, a frozen fusion underlay supplies pixels left uncovered by the model, while baselines are scored on their native outputs. The dynamic-region scores use the no-fusion evaluation. The full-frame comparison therefore combines the model's output with an underlay in a way the baseline scores do not.
A work in progress with clear limits
The model's object-by-object world model also has explicit boundaries. It assumes no collisions, cannot instantiate objects that appear after time t, and lacks shape completion for object sides that were not observed. Those limits leave open important cases for a dynamic scene representation.
The static bank cannot synthesize background that becomes newly visible, and the paper evaluates prediction only at short horizons, leaving longer-horizon rollout open. The manuscript is labeled a preprint and describes the work as in progress. Its conclusions are confined to the reported KITTI-MOT future-prediction and past-reconstruction protocols.
Paper data and sources
Original title: 4DGS-WAM: Bridging Past and Future with an Object-Centric World Action Model based on 4D Gaussian Splatting
Authors: Yueen Ma, Zenglin Xu, Irwin King
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text