Triangle Splatting represents scenes using semi-transparent triangles, but it evaluates each triangle at the pixel center and culls any triangle with an inradius smaller than one pixel. The former causes severe aliasing, while the latter culls most of the scene at lower resolutions. We show that the integral of the Triangle Splatting window over a pixel admits an exact closed-form solution. The angle bisectors split the triangle into three sub-triangles, in each of which the window is a power of the distance to one edge. Each piece therefore has a closed-form antiderivative, and Green's theorem reduces the integral and its gradients to one term per side of the sub-triangle clipped against the pixel. The integral over the whole triangle, which we show has a closed form, then allows conservative culling, removing only triangles that cannot meet the renderer's alpha cutoff at any pixel. This exact formulation anti-aliases existing models without retraining, exhibits less temporal aliasing than 2x2 supersampling, and enables multi-scale training, matching or outperforming supersampled training at reduced resolutions.
Figures & tables
Figure 2: Culling sub-pixel triangles. Triangle Splatting at (a) the training resolution and (b) \nicefrac18 of it. Left half of each image: without culling and the right half with the released culling of triangles smaller than a pixel. At \nicefrac18 , many triangles are smaller than a pixel, so culling removes nearly the entire scene, which severely impacts PSNR (27.2 dB → 18.0 dB on this scene).
Figure 3: Pipeline per triangle. (i) Given a triangle T and a pixel Π , we aim to compute the integral of the window function inside the pixel. (ii) The bisectors split the triangle into sub-triangles W0,W1,W2 , in each of which the window depends only on the distance to one edge. (iii) Each sub-triangle is clipped to the pixel, giving Di=Π∩Wi . (iv) Green’s theorem turns the integral over Di into one closed-form term per side ei .
Figure 4: Zooming out after single-scale training on three Mip-NeRF 360 scenes. Each crop shows the training resolution (top) and its \nicefrac18 (bottom), upsampled. The left shows triangle-based methods, the middle shows Gaussian anti-aliasing methods, and the right shows ground truth. Point sampling aliases at \nicefrac18 ; 2×2 supersampling reduces this but still leaves some aliasing, e.g., on bonsai’s tablecloth.
Method
1
\nicefrac12
\nicefrac14
\nicefrac18
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Mip-NeRF 360
Triangle Splatting no culling
27.16
0.815
0.190
27.78
0.850
0.121
26.87
0.827
0.130
24.83
0.765
0.178
Triangle Splatting supersampled
27.27
0.819
0.192
28.41
0.868
0.113
29.69
0.904
0.072
29.90
0.908
0.072
Ours
27.26
0.819
0.187
28.36
0.866
0.113
29.33
0.902
0.075
29.26
0.920
0.062
Tanks and Temples
Table 1: Zoom-out after single-scale training. Test views are rendered at 1 to \nicefrac18 of the training resolution and compared with the box-downsampled ground truth, and we report the mean over the scenes of each data set. Triangle Splatting no culling is the released method without culling of sub-pixel triangles in training and rendering, as the cull breaks down when zooming out (supplementary material), and Triangle Splatting supersampled renders that model with 2×2 samples per pixel. All models are trained in our framework using the released settings and a numerically stable opacity update for cloned triangles (see supplementary material). Blue marks the best value.
Mip-NeRF 360
Tanks and Temples
Deep Blending
Renderer
\nicefrac14
\nicefrac18
\nicefrac14
\nicefrac18
\nicefrac14
\nicefrac18
Point sampling
10.1
20.7
13.1
27.0
3.7
8.6
2×2 supersampling
3.8
9.2
5.1
13.0
1.4
3.7
Ours
1.9
3.3
2.7
4.5
0.9
1.7
Table 2: Temporal stability. Error of the frame-to-frame differences ( ×10−3 , lower is better) against a 16×16 supersampled render of the same model, at \nicefrac14 and \nicefrac18 of the training resolution. All renderers draw the Triangle Splatting no culling model of Tab. 1 . We report the mean over all scenes. Blue marks the best value.
Method
1
\nicefrac12
\nicefrac14
\nicefrac18
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Mip-NeRF 360
Triangle Splatting no culling
25.69
0.735
0.293
26.87
0.815
0.170
27.82
0.864
0.106
26.45
0.835
0.130
Triangle Splatting supersampled
26.65
0.784
0.236
27.98
0.854
0.133
29.56
0.905
0.074
30.27
0.921
0.063
Ours
26.91
0.802
0.218
28.29
0.865
0.121
29.93
0.913
0.068
31.00
0.941
0.046
Tanks and Temples
Table 3: Zoom-out after multi-scale training with the scenes of Tab. 1 , and on NeRF Synthetic at the four scales of its multi-scale version [ 2 ] . Each iteration renders a training view at a random scale from 1 to \nicefrac18 , with full resolution forced in 30% of the iterations as in Mip-Splatting [ 36 ] , and every model is trained with the renderer it is tested with. Triangle Splatting no culling has no NeRF Synthetic row, since its training ran out of memory on lego, and Triangle Splatting supersampled is trained and rendered with 2×2 samples per pixel. Blue marks the best value.
Renderer
1
\nicefrac12
\nicefrac14
\nicefrac18
Triangle Splatting no culling
202
218
143
69
+ depth split, 8×8 tiles
197
287
357
397
2×2 supersampling
108
204
220
144
+ depth split, 8×8 tiles
101
187
280
350
Ours, exact
81
97
56
27
+ depth split, 8×8 tiles
99
164
238
322
Table 4: Inference speed in frames per second (FPS) of the single-scale models of Tab. 1 , mean over the nine Mip-NeRF 360 scenes (RTX 5090, median over test views). 2×2 supersampling draws the Triangle Splatting no culling model.
Mip-NeRF 360
Tanks and Temples
Method
1
\nicefrac18
1
\nicefrac18
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
3DGS [ 18 ]
27.52
0.814
0.217
18.79
0.567
0.240
23.87
0.851
0.170
17.10
0.619
0.207
Mip-Splatting [ 36 ]
27.66
0.827
0.190
28.89
0.925
0.060
23.79
0.858
0.157
24.07
0.919
0.062
Analytic-Splatting [ 22 ]
27.52
0.816
0.215
28.47
0.917
0.074
23.75
0.853
0.173
23.78
0.912
0.067
Gaussian Blending [ 20 ]
26.86
0.803
0.237
30.37
0.931
0.068
23.27
0.846
0.188
25.41
0.936
0.053
Table 5: Comparison with Gaussian-based methods after single-scale training, at the training resolution ( 1 ) and at \nicefrac18 of it, with the scenes and resolutions of Tab. 1 . The Gaussian-based methods are trained with the authors’ code and settings and evaluated with the same protocol as ours. All scales, data sets and multi-scale training are in the supplementary material. Blue marks the best value.
Point sampling / Ours
Samples
1
\nicefrac12
\nicefrac14
\nicefrac18
1×1
42.6 / 52.2
36.3 / 45.7
30.2 / 39.5
26.1 / 34.5
2×2
52.6 / 58.8
47.8 / 54.7
42.4 / 49.2
36.2 / 43.3
4×4
60.1 / 63.7
56.5 / 60.6
52.8 / 56.8
48.2 / 52.0
8×8
66.4 / 68.0
63.2 / 65.1
60.0 / 61.8
56.7 / 58.1
Table 6: Where the remaining error comes from. Our single-scale Mip-NeRF 360 models rendered with k×k samples per pixel: point samples, blended per pixel, or our exact pixel integral per sub-pixel, blended with one α per triangle and sub-pixel. PSNR (dB) against the 16×16 point-sampled render of the same model that agrees with 8×8 to >56 dB, indicating convergence. Means over the nine scenes.
Rendered with
1
\nicefrac12
\nicefrac14
\nicefrac18
Point sampling, with cull
27.18
26.76
20.13
12.93
Point sampling, without cull
27.04
27.65
26.83
24.86
2×2 supersampling
27.18
28.32
29.59
29.86
Ours
27.18
28.25
29.19
29.28
Table 7: Drop-in renderer. PSNR (dB) on Mip-NeRF 360 of the single-scale Triangle Splatting original models, trained with point sampling and the inradius cull [ 16 ] , rendered without retraining.
Figure 5: Culling threshold sweep on the single-scale Mip-NeRF 360 models, with depth split, mean over the nine scenes. The threshold is given in multiples of the alpha cutoff \nicefrac1255 , off disables the cull, and the dashed line marks ours. Top left: PSNR change against no culling. Top right: frame rate. Bottom left: share of culled triangles.
Figure 6: Spatio-temporal slices of a sub-pixel pan at \nicefrac18 of the training resolution. The camera moves by \nicefrac116 pixel per frame, and each slice stacks the marked image row over the frames. All renderers draw the same model using Triangle Splatting no culling . The reference renders it with 16×16 supersampling, since no captured images exist for the pan. Point sampling breaks thin structures into dashes as they enter and leave the sample positions, 2×2 supersampling reduces this, and ours stays closest to the reference.
Figure 7: Error comparison. We render our single-scale Mip-NeRF 360 models at \nicefrac18 of the training resolution and measure the error against a 16×16 point-sampled render of the same model, as in Tab. 6 . The second and third columns show the per-pixel absolute error of 2×2 point samples and of ours with one α per pixel, with their PSNR in dB. The last column shows which method has the higher error, red for ours and blue for 2×2 point samples. Our errors concentrate on edges, whereas point sampling mainly errs on textured areas, which agrees with per-pixel compositing being the main source of error in our method.
Method
1
\nicefrac12
\nicefrac14
\nicefrac18
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Mip-NeRF 360
3DGS [ 18 ]
27.52
0.814
0.217
25.09
0.807
0.180
21.29
0.690
0.208
18.79
0.567
0.240
Mip-Splatting [ 36 ]
27.66
0.827
0.190
28.81
0.877
0.110
29.54
0.913
0.070
28.89
0.925
0.060
Analytic-Splatting [ 22 ]
27.52
0.816
0.215
28.64
0.866
0.137
29.18
0.903
0.091
28.47
0.917
0.074
Gaussian Blending [ 20 ]
26.86
0.803
0.237
27.99
0.853
0.161
29.36
0.900
0.104
30.37
0.931
0.068
Table 8: Zoom-out after single-scale training on Mip-NeRF 360 [ 3 ] , Tanks and Temples [ 19 ] (truck, train), Deep Blending [ 14 ] (drjohnson, playroom) and NeRF Synthetic [ 27 ] , for every method. Test views are rendered at 1 to \nicefrac18 of the training resolution and compared with the box-downsampled ground truth. We report the mean across scenes for each data set. Triangle Splatting original point-samples and culls sub-pixel triangles, as released. Triangle Splatting no culling keeps them in training and rendering, and Triangle Splatting supersampled renders that model with 2×2 samples per pixel. The triangle models are trained in our framework using the released settings and the stable opacity update from Sec. 9 , while the Gaussian methods are trained using the authors’ code and settings. All methods are evaluated with the same protocol. Blue marks the best value.
Method
1
\nicefrac12
\nicefrac14
\nicefrac18
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Mip-NeRF 360
Mip-Splatting [ 36 ]
27.52
0.819
0.208
28.82
0.876
0.117
30.36
0.922
0.065
31.31
0.948
0.042
Analytic-Splatting [ 22 ]
27.49
0.810
0.227
28.82
0.868
0.136
30.37
0.919
0.076
31.35
0.948
0.046
Gaussian Blending [ 20 ]
26.95
0.800
0.243
28.17
0.858
0.153
29.75
0.912
0.088
31.11
0.945
0.052
Triangle Splatting original
26.01
0.778
0.247
26.91
0.828
0.160
27.65
0.857
0.116
23.28
0.729
0.219
Table 9: Zoom-out after multi-scale training on Mip-NeRF 360 [ 3 ] , Tanks and Temples [ 19 ] and Deep Blending [ 14 ] , with the scenes and resolutions of Tab. 1 , and on NeRF Synthetic [ 27 ] (8 scenes) at the four scales of its multi-scale version [ 2 ] , whose lower resolutions are box-downsampled renders. Training follows Tab. 3 . Triangle Splatting original point-samples and culls sub-pixel triangles, and we train it at all scales with our code, since the released code has no multi-scale training. Triangle Splatting no culling keeps them in training and rendering, and is missing on NeRF Synthetic, since its training ran out of memory on lego. Triangle Splatting supersampled is trained and rendered with 2×2 samples per pixel. Analytic-Splatting and Gaussian Blending are trained with the authors’ multi-scale code. The released Mip-Splatting code only trains at several scales on NeRF Synthetic, so on the other data sets we load every view at all four scales, as Analytic-Splatting does. All methods are evaluated using the same protocol, and we report the mean across scenes for each data set. Blue marks the best value.
Feed-forward 3D Gaussian Splatting (FF-3DGS) emerges as a fast and robust solution for sparse-view 3D reconstruction and novel view synthesis (NVS). However, existing FF-3DGS methods are built on incorrect screen-space dilation filters, causing severe rendering artifacts when rendering at out-of-distribution sampling rates. We firstly propose an FF-3DGS model, called AA-Splat, to enable robust anti-aliased rendering at any resolution. AA-Splat utilizes an opacity-balanced band-limiting (OBBL) design, which combines two components: a 3D band-limiting post-filter integrates multi-view maximal frequency bounds into the feed-forward reconstruction pipeline, effectively band-limiting the resulting 3D scene representations and eliminating degenerate Gaussians; an Opacity Balancing (OB) to seamlessly integrate all pixel-aligned Gaussian primitives into the rendering process, compensating for the increased overlap between expanded Gaussian primitives. AA-Splat demonstrates drastic improvements with average 5.4∼7.5dB PSNR gains on NVS performance over a state-of-the-art (SOTA) baseline, DepthSplat, at all resolutions, between 4× and 1/4×. Code will be made available.
Taewoo Suh, Sungpyo Kim, Jongmin Park +1
Korea Advanced Institute of Science and Technology
Rendering 3D scenes as pixel art requires that discrete pixels remain stable as the camera moves. Existing methods snap the camera to a grid. Under orthographic projection, this works: every pixel shifts by the same amount, and a single snap corrects all of them. Perspective breaks this. Pixels at different depths drift at different rates, and no single snap corrects all depths. Texel splatting avoids this entirely. Scene geometry is rendered into a cubemap from a fixed point in the world, and each texel is splatted to the screen as a world-space quad. Cubemap indexing gives rotation invariance. Grid-snapping the origin gives translation invariance. The primary limitation is that a fixed origin cannot see all geometry; disocclusion at probe boundaries remains an open tradeoff.
Recent radiance field methods represent scenes with 2D primitives that offer surface alignment and efficient rasterization, from Gaussian disks to triangles, yet all rely on convex boundaries: curved and concave structures demand excessive primitives. We introduce Deformable Triangle Splatting, which augments each triangle with K control points per edge, each parameterized by a single learnable scalar displacement that shifts the boundary inward or outward, enabling non-convex shape representation while preserving the three base vertices that define the 3D plane. To render these non-convex primitives differentiably, we design a rasterization pipeline in the triangle's barycentric coordinate space, ensuring view-consistent rendering. A winding number test determines whether each pixel lies inside the deformed primitive, and a window function controlled by two learnable parameters, sharpness and corner smoothness, together with a per-primitive scalar opacity, produces the smooth opacity transition from interior to boundary. Validation is done in a variety of real-world scenes, outperforming recent works based on non-volumetric primitives in terms of visual quality and versatility while still achieving competitive rendering efficiency.
Oriol Jiménez-Ayguadé, Antonio Agudo
Institut de Robòtica i Informàtica Industrial, CSIC-UPC, Barcelona, Spain