TOM-GS: Editable Video Representation via Temporal Opacity Modulation of Static 3D Gaussians
Marek Lisowski, Łukasz Smoliński, Kornel Howil, Piotr Biliński, Marcin Mazur, Przemysław Spurek
cs.CV
2026-07-22
TOM-GS keeps 3D Gaussians static and models motion only through a learned temporal opacity, leading DAVIS fidelity while remaining a standard, directly-editable 3DGS asset.
Turning a monocular video into an editable 3D representation has been stuck on a trade-off. Implicit neural representations are black boxes you cannot edit. Dynamic 3D Gaussian Splatting either drives Gaussians with deformation fields or uses specialized folded distributions (VeGaS) to model nonlinear motion. These capture complex dynamics, but the geometry is no longer standard 3DGS, so off-the-shelf 3D editors and physics engines cannot touch it. Expressiveness and editability have been an either-or. TOM-GS picks editability.
The key decision is to keep geometry fully static. TOM-GS uses ordinary 3D Gaussians whose position, rotation, scale, and color are pinned to one world frame and never move.
Motion comes entirely from opacity over time. Each Gaussian gets a learnable temporal mean and scale, and its opacity follows a Gaussian curve in time: it fades in when it should appear and out when it should leave. Objects appear to move because a string of static Gaussians laid out along the trajectory light up one after another. The whole scene is a static point cloud where each point has its own time window.
The supporting choices all protect the static geometry:
The editing pipeline converts the Gaussian cloud into a tetrahedral mesh (following GaMeS), applies edits on the tetrahedra (manual scale, duplicate, remove, or a soft-body simulation in Blender), and converts back.
The cost is expressiveness: motion is approximated by opacity flicker rather than true trajectory deformation.
On the DAVIS benchmark, TOM-GS leads the editable-method family on quality.
| Method | Avg PSNR | Avg SSIM |
| TOM-GS | 36.50 | 0.98 |
| VeGaS | 33.31 | 0.94 |
| Splatter-a-Video | 28.44 | 0.84 |
Video-level metrics reach PSNR 39.80 and SSIM 0.98, ahead of GaussianVideo, VeGaS, and others (LPIPS 0.030 is slightly worse than GaussianVideo's 0.021). Against neural representations such as the NeRV family, the average PSNR of 35.06 also beats VeGaS's 32.42.
Ablations make the method legible. Remove temporal opacity (plain static 3DGS) and only 24.36 remains, so temporal opacity is the whole game. Replacing it with a neural opacity predictor gives 29.19, so an explicit per-Gaussian window wins. Efficiency averages 8.51M Gaussians, 90 minutes of training, and 63.4 FPS rendering.
This is the first editable video representation that both leads on fidelity and is a stock 3DGS asset. Because the geometry never deforms, the result is ordinary 3D Gaussians that drop straight into tools like Blender for scaling, deletion, or soft-body physics. Prior editable methods either could not plug into the toolchain or trailed on quality. For anyone doing 3D video editing or physics simulation, the output is no longer a view-only reconstruction; it is an asset you can actually edit.
The authors name two. Long videos and abrupt scene changes are hard, because each Gaussian has only one time window and cannot represent a point that appears, vanishes, then reappears. The Gaussian count also grows (capped at 10M), so memory and inference cost scale with video length. On top of that, the method depends on AnyCam pose quality and simplifies foreground editing to single-plane geometry. The 10M cap is itself a quality-for-memory compromise; lifting it buys more quality.