TruncGradGS: Gauss Splatting Method
- TruncGradGS are enhancements for 3D Gaussian Splatting that modify gradients and rasterization coverage to prevent gradient vanishing in spatial tails of Gaussian primitives, leading the improved optimization for low-opacity or dynamic Gaussians.
- The method applies a continuous, piecewise linear surrogate for Gaussian density derivatives outside a prescribed density isocontour, expanding rasterization coverage and applying adaptive corrections to optimization, improving results in both static and dynamic settings.
- Notably, TruncGradGS 2D applications improve PSNR from 18.80 to 20.70 and SSIM form 0.4960 to0.6624 on random-initialized static data, and for Neural 3D Video, increasing PSNR from 30.22 to 31.70, SSIM from 0.9443 to .9493,.
TruncGradGS is a gradient-modification method for 3D Gaussian Splatting (3DGS) that addresses gradient vanishing in the spatial tails of projected Gaussian primitives. Its central operation replaces exponentially small Gaussian-density derivatives outside a prescribed density isocontour with a continuous, piecewise linear surrogate, while retaining the ordinary derivative in the high-density interior. The method also expands rasterization coverage for low-opacity “dead” Gaussians, applies sign-based filtering, interleaves exploratory truncated-gradient updates with conventional Adaptive Density Control, and delays pruning. These mechanisms are designed to provide long-range optimization signals for poorly initialized or dynamically displaced primitives (Morales et al., 3 Sep 2026).
1. Problem setting and gradient vanishing
3DGS represents a scene using anisotropic three-dimensional Gaussian primitives. A primitive is parameterized by a three-dimensional mean , covariance , opacity , and view-dependent appearance coefficients, typically spherical harmonics. After projection through the camera, each primitive becomes a two-dimensional elliptical splat.
For a pixel , the projected Gaussian density is
where the projected covariance is approximated by
Here, is the viewing transformation and is the Jacobian of the affine approximation to the projective camera transformation. The rendered pixel color is obtained by alpha compositing sorted Gaussians:
Let
For a two-dimensional projected Gaussian with conic coefficients 0, the spatial derivatives have the form
1
and
2
Although the linear displacement factors increase with distance, the Gaussian density decays exponentially. Consequently,
3
as the pixel moves into the Gaussian tail. A distant pixel can therefore exhibit a substantial reconstruction error while producing almost no positional update for a Gaussian that is currently far from that pixel.
The resulting issue primarily affects spatial attributes: projected and three-dimensional means, primitive positions, and indirectly scales and covariances. Opacity and appearance parameters, including spherical-harmonic coefficients, can also receive weak or absent contributions when a primitive does not participate meaningfully in the rasterization path.
2. Rasterization locality and ineffective support
The optimization problem is intensified by the efficiency mechanisms used in Gaussian rasterization. Images are divided into 4 tiles, and each Gaussian is assigned only to tiles intersecting its projected bounding box. Pixels outside those tiles cannot contribute to its backward pass.
Within an assigned tile, contributions whose scalar opacity-density product is below a rasterizer threshold are discarded. In simplified form, the filtering condition is
5
Thus, although a Gaussian has mathematical support over the entire image plane, its effective computational support is limited by tile assignment and density filtering. A badly initialized primitive, a primitive displaced by dynamic motion, or a primitive assigned to the wrong region may receive no useful gradient from the pixels that would indicate where it should move.
Traditional 3DGS addresses inadequate coverage partly through cloning, splitting, scale expansion, and adaptive density control. Increasing a Gaussian’s radius, however, does not eliminate the exponential decay of its derivative in the far tail. TruncGradGS instead modifies the backward derivative and expands the rasterization coverage of selected low-opacity Gaussians.
The method is not a replacement for the forward renderer. Forward images continue to use ordinary Gaussian splatting. The modification is confined to the backward computation and affects the parameter update
6
where 7 is the deliberately modified gradient field.
3. Piecewise truncated-gradient formulation
TruncGradGS defines an isocontour using a density threshold 8. The squared Mahalanobis distance from a query pixel to the projected Gaussian mean is
9
The truncation boundary is characterized by
0
or equivalently,
1
Inside this contour, the exact Gaussian derivative is retained. Outside it, the derivative is replaced by a continuous linear surrogate. For the 2 coordinate, the rule is
3
The same construction is applied to the 4 coordinate. The variables have the following roles:
- 5 is the projected Gaussian density at the query point.
- 6 defines the truncation contour.
- 7 is the boundary point on the 8-isocontour.
- 9 is the user-defined slope of the exterior surrogate.
- 0 is the intercept chosen for continuity.
- 1 are the conic coefficients of the projected Gaussian.
The symbol 2 is overloaded: it denotes both a conic coefficient and the intercept of the surrogate line; these are distinct quantities.
Interior and exterior regions
When
3
the query point lies inside the threshold contour, equivalently 4, and ordinary differentiation is used.
Outside the contour, the exponentially small derivative is replaced by a line anchored at the boundary. If the boundary derivative is negative, the surrogate is based on 5 and the minimum is selected. If the boundary derivative is positive, it is based on 6 and the maximum is selected. This preserves the derivative’s directional behavior while preventing its magnitude from collapsing in the tail.
Continuity at the boundary
The surrogate satisfies
7
which gives
8
The resulting piecewise field is continuous at the truncation boundary. The slope 9 controls the strength and range of the long-distance signal: a smaller value produces a flatter exterior gradient, whereas a larger value produces a sharper decay away from the boundary.
Boundary construction
The boundary point is parameterized along the ray from the query pixel to the Gaussian mean:
0
It satisfies
1
Solving the resulting quadratic gives
2
where
3
4. Rasterization integration and update policy
A truncated derivative is useful only if the corresponding Gaussian-pixel pair reaches the backward computation. TruncGradGS therefore pads the tile-coverage radius of selected low-opacity Gaussians.
A dead Gaussian is defined as a primitive whose opacity is below the pruning threshold,
4
For these primitives, the rasterization radius is expanded so that more image tiles are assigned to them. Distant pixels can then participate in the backward computation even when the ordinary projected support is too small.
Radius padding and truncated-gradient computation are not applied indiscriminately. Active Gaussians use ordinary gradients. The truncated mechanism is reserved for dead Gaussians because applying the surrogate to already well-positioned primitives can bias otherwise useful local solutions and destabilize convergence.
The method uses three gradient paths:
- Normal path: the Gaussian passes the density filter and receives the ordinary gradient.
- Truncated path: the Gaussian has sufficiently low density to be ineffective in the ordinary rasterizer, but receives the surrogate long-range gradient.
- Skip path: the Gaussian is discarded and is not eligible for the truncated signal.
The method additionally applies sign filtering. Truncated gradients are used only when
5
This condition retains long-range attraction when the Gaussian’s contribution can reduce the loss. Without this filtering, the surrogate can produce both long-range attraction and repulsion, causing Gaussians to move away from useful regions or repel one another.
The training schedule interleaves truncated-gradient optimization with conventional optimization using Adaptive Density Control (ADC):
- perform truncated-gradient updates for eligible dead Gaussians;
- perform normal optimization with ADC enabled;
- delay pruning until the end of training.
Delayed pruning allows low-opacity primitives to receive exploratory positional updates before being removed. It also prevents truncated view-space gradients from contaminating ADC statistics.
5. Static and dynamic training pipelines
For static scenes, the method begins with either random sparse Gaussian initialization or COLMAP/SfM point-cloud initialization. Gaussian means, scales, rotations, opacities, and spherical-harmonic coefficients are optimized through differentiable rendering. Active and interior contributions use ordinary gradients, whereas eligible dead and far-field contributions use truncated gradients. The original densification and splitting machinery is retained, truncated and normal phases are interleaved, and pruning is delayed.
At inference time, the ordinary renderer is used. Radius padding and truncated backward computation therefore do not affect rendering speed.
For dynamic Gaussian Splatting, the same modification is inserted into the backward pass of dynamic representations. The method is demonstrated with 4DGS, CEM-4DGS, and 4D-Scaffold. Dynamic representations are initialized from the first frame and optimized sequentially over time. Because temporal motion can displace primitives from the regions they should explain, truncated gradients provide long-range spatial signals for primitives that become poorly aligned with later frames.
To reduce dynamic-scene overhead, the method constructs viewpoint-dependent motion masks using frame differences and image processing. Static pixels are ignored during the special truncated-gradient computation because they do not contribute to the dynamic motion problem.
6. Empirical evaluation
The evaluation reports PSNR, SSIM, LPIPS, training time, and final Gaussian count. Static experiments use Mip-NeRF360 and static frames from the proposed synthetic dynamic benchmark. Dynamic experiments use Neural 3D Video and a synthetic benchmark containing the scenes alley, windy tree, water cup, neon city, bouncy balls, and underwater. The synthetic benchmark contains 300 frames per scene, 10-second sequences, 30 fps, 25–45 synchronized cameras, 6 resolution, Blender-generated path-traced images, and one to four held-out test viewpoints.
Static reconstruction
Under random initialization, TruncGradGS improves the reported metrics for both the proposed static data and Mip-NeRF360. For 3DGS on the proposed static data, the method changes LPIPS from 7 to 8, SSIM from 9 to 0, and PSNR from 1 to 2, while reducing the final Gaussian count from 3M to 4M. On Mip-NeRF360 with random initialization, LPIPS changes from 5 to 6, SSIM from 7 to 8, and PSNR from 9 to 0.
The method also transfers to 2DGS. For random initialization on the proposed static data, 2DGS + TruncGradGS improves PSNR from 1 to 2 and SSIM from 3 to 4. Under COLMAP initialization, the gains are smaller and can be mixed. On Mip-NeRF360 with COLMAP-initialized 3DGS, PSNR improves from 5 to 6, while LPIPS changes from 7 to 8 and SSIM from 9 to 0.
Training is substantially slower. For random-initialized 3DGS on the proposed static data, training time increases from 14 minutes to 55 minutes. On Mip-NeRF360 it increases from 27 to 56 minutes. With COLMAP initialization, the corresponding times are 9 to 43 minutes on the proposed data and 21 to 86 minutes on Mip-NeRF360.
Dynamic reconstruction
On Neural 3D Video, TruncGradGS improves 3DGS PSNR from 1 to 2, SSIM from 3 to 4, and LPIPS from 5 to 6, while reducing the Gaussian count from 7M to 8M.
For 4DGS, PSNR increases from 9 to 0 dB and SSIM from 1 to 2, although LPIPS changes from 3 to 4. For 4D-Scaffold, PSNR increases from 5 to 6 dB and the Gaussian count decreases from 7M to 8M. For CEM-4DGS, LPIPS improves from 9 to 0, while PSNR changes from 1 to 2.
On the proposed synthetic benchmark, the strongest result is reported for dense-COLMAP-initialized 4DGS. TruncGradGS changes LPIPS from 3 to 4, SSIM from 5 to 6, and PSNR from 7 to 8 dB, while reducing the Gaussian count from 9M to 00M.
Ablation findings
The ablation study evaluates the alley and tree scenes. Removing the truncated gradient degrades the principal mechanism. Removing radius padding limits the tiles through which long-range gradients can be propagated. Removing delayed pruning prevents low-opacity Gaussians from receiving exploratory updates. Removing dead-Gaussian filtering is particularly unstable: on the alley scene, LPIPS rises to 01, SSIM falls to 02, and PSNR falls to 03 dB.
The full method is not best on every individual scalar metric. For example, disabling negative-alpha filtering produces slightly higher alley SSIM and PSNR, and disabling delayed pruning produces higher tree SSIM and PSNR. These variants, however, are less reliable conceptually and can cause repulsive behavior or instability.
7. Limitations and interpretation
TruncGradGS is not ordinary gradient descent on the exact forward renderer everywhere. It retains exact local derivatives but introduces a biased, directionally guided surrogate in the Gaussian tail. Its effectiveness therefore depends on the choice of density threshold, surrogate slope, dead-Gaussian criterion, sign filter, radius padding, and pruning schedule.
The principal computational cost is increased Gaussian-pixel participation during backward propagation. The reported training slowdown ranges from approximately 04 to 05 times the baseline, depending on scene complexity and primitive count. Temporary memory pressure can also increase because padded dead Gaussians are assigned to more tiles. No separate peak-memory table is reported.
Potential divergence arises when the surrogate gradient dominates the true gradient, either for a poorly positioned Gaussian receiving a strong exploratory signal or for a well-fitted Gaussian whose exact gradient is already near zero. The method addresses this through dead-Gaussian filtering, negative-alpha filtering, interleaved optimization, and delayed pruning. A poorly selected slope 06 can produce weak exploration or excessive movement, although the supplied evaluation does not provide a numerical sensitivity sweep.
The strongest improvements occur under poor initialization, weakly constrained geometry, or substantial temporal displacement. Improvements are less consistent when COLMAP provides a strong initialization. The proposed dynamic benchmark is synthetic and therefore does not replace real-world captures; its motion, rendering, and camera statistics may differ from physical acquisition.
Conceptually, TruncGradGS treats Gaussian Splatting optimization as a support-coverage problem. A primitive’s useful optimization support is restricted not only by the mathematical Gaussian tail but also by tile culling, density filtering, and opacity-based pruning. TruncGradGS expands this effective support in the backward pass: exact derivatives govern local refinement, while continuous surrogate derivatives provide long-range attraction for selected dead or weakly contributing Gaussians. This makes gradient propagation complementary to initialization, densification, pruning, and deformation modeling rather than a replacement for them.