Triangle-based splatting connects novel view synthesis with conventional graphics pipelines, but often requires excessive tessellation to preserve appearance details. We present T2-Splat, which decouples topology adaptation from appearance refinement. Safe Restricted Delaunay Triangulation (Safe RDT) initializes mesh connectivity. Dynamic Topology Evolution (DTE) then adapts triangle allocation through edge collapse and vertex split operations. Residual per-face Mip-textures then refine appearance on the fixed mesh using scale-adaptive filtering. Experiments on Mip-NeRF 360, Tanks & Temples, and DTU demonstrate competitive results with compact meshes. T2-Splat improves PSNR by 0.71dB over MeshSplatting on Mip-NeRF 360. Additional compression experiments demonstrate flexible control of mesh complexity, supporting face budgets as low as 10% of the original face count. Project page: https://t2-splat.github.io/.
Figures & tables
Figure 1: Top: our method better preserves high-frequency details (e.g., text) than MeshSplatting Held et al. [2025a] and MiLo Guédon et al. [2025] on the shelf scene (STRinGS-360 Raundhal et al. [2026] ). Bottom: on the garden scene, T 2 -Splat achieves the highest PSNR (25.68 dB) among the compared methods using only 2.7M faces, versus 92M for GOF Yu et al. [2024] and 15M for MiLo Guédon et al. [2025] .
Figure 2: T2 -Splat is an end-to-end differentiable rendering framework that bridges high-fidelity novel view synthesis and standard graphics pipelines. In Stage I, it constructs a manifold mesh from an unstructured triangle soup via Safe Restricted Delaunay Triangulation (Safe RDT), and introduces Dynamic Topology Evolution (DTE) to adaptively optimize topology using QEM-based edge collapses and gradient-driven vertex splits. Stage II employs a scale-aware residual Mip-texture model with trilinear sampling, decoupling appearance from geometry to ensure high-quality, anti-aliased rendering even under extreme mesh compression.
Figure 3: Qualitative results. Compared to baselines that suffer from blurry textures or geometric distortions, our T 2 -Splat utilizes per-face textures to effectively eliminate texture distortion and preserves significantly sharper high-frequency details (e.g., fine text and intricate surface patterns) and maintains better structural integrity.
Method
Characteristics
Mip-NeRF360 dataset
Tanks & Temples
Mesh
Color
Connect
PSNR ↑
LPIPS ↓
SSIM ↑
#Verts ↓
PSNR ↑
LPIPS ↓
SSIM ↑
#Verts ↓
3DGS
–
–
–
27.21
0.214
0.815
–
23.14
0.183
0.841
–
Triangle Splatting
–
–
–
27.16
0.191
0.814
–
23.14
0.143
0.857
–
2DGS
✗
✗
✓
15.36
0.474
0.498
2M
14.23
0.485
0.569
16M
GOF
✗
✗
✓
20.78
0.465
0.573
33M
21.69
0.326
0.690
12M
RaDe-GS
✗
✗
✓
23.56
0.361
0.668
31M
20.51
0.344
0.659
10M
Table 1: Mesh-based NVS results on Mip-NeRF360 and Tanks & Temples. We report PSNR, SSIM, LPIPS, and vertex count. Mesh , Color , and Connect indicate whether the method directly outputs a mesh, whether the mesh carries color without post-processing, and whether the output is connected, respectively.
Method
24
37
40
55
63
65
69
83
97
105
106
110
114
118
122
Mean
3DGS
2.14
1.53
2.08
1.68
3.49
2.21
1.43
2.07
2.22
1.75
1.79
2.55
1.53
1.52
1.50
1.96
2DGS
0.48
0.91
0.39
0.39
1.01
0.83
0.81
1.36
1.27
0.76
0.70
1.40
0.40
0.76
0.52
0.80
GOF
0.50
0.82
0.37
0.37
1.12
0.74
0.73
1.18
1.29
0.68
0.77
0.90
0.42
0.66
0.49
0.74
RaDe-GS
0.46
0.73
0.33
0.38
0.79
0.75
0.76
1.19
1.22
0.62
0.70
0.78
0.36
0.68
0.47
0.68
MiLo
0.43
0.74
0.34
0.37
0.80
0.74
0.70
1.21
1.22
0.66
0.62
0.80
0.37
0.76
0.48
0.68
Triangle Splatting
0.98
1.07
1.07
0.51
1.67
1.44
1.17
1.32
1.75
0.98
0.96
1.11
0.56
0.93
0.72
1.08
Table 2: Chamfer distance on the DTU dataset. We report Chamfer distance across 15 scenes. T 2 -Splat achieves a joint second-best mean Chamfer distance of 0.68 and the lowest error on three scenes (63, 83, and 105).
Table 6
Safe RDT
DTE
MipTex
PSNR ↑
LPIPS ↓
SSIM ↑
#Faces ↓
Time (min) ↓
(a)
–
–
–
24.66
0.30
0.734
5.3M
54
(b)
✓
–
–
24.81
0.31
0.740
3.0M
43
(c)
✓
✓
–
25.12
0.29
0.756
2.5M
57
(d)
✓
–
4×4
25.20
0.27
0.762
3.0M
65
(e)
✓
–
8×8
25.31
0.24
0.769
3.0M
81
(f)
✓
✓
4×4
25.43
0.26
0.775
2.5M
72
Table 5: Ablations on Mip-NeRF 360. We progressively enable Safe RDT, Dynamic Topology Evolution (DTE), and residual per-face Mip-textures (MipTex) from baseline (a). DTE improves rendering quality with fewer faces, while Mip-textures further refine appearance on fixed geometry. Results are averaged equally over nine scenes.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Outdoor
Indoor
feature learning rate
0.0016
0.004
opacity learning rate
0.03
0.05
vertex pos. learning rate
0.0015
0.0015
texture learning rate
0.0005
0.0005
densification start
500
500
densification end
10000
10000
Appendix
Table 6: Main hyperparameters used in our experiments.
Iteration / event
Representation update
Optimization role
500–10,000
Densification every 500 iterations
Establish geometric support before mesh construction; pruning starts at 4,000 outdoors and 3,000 indoors.
11,000
Safe RDT
Replace soup connectivity with a constrained face set; retain vertex parameters and optimizer states.
12,000–20,000
DTE every 1,000 iterations
Adapt connectivity using edge collapse and gradient-driven refinement while optimizing geometry and base appearance.
20,000–30,000
DTE interval completed
Continue Stage I optimization of vertex positions, opacities, and SH coefficients.
30,000–35,000
Fixed topology, positions, opacity
Jointly optimize vertex SH coefficients and residual per-face textures.
Appendix
Table 7: Training stages and their roles. Iteration indices describe the reported schedule; the beginning and end of each interval mark stage transitions.
Foreground only
Full scene extraction
Scene
2DGS
Gaussian Wrapping
GOF
RaDe-GS
MiLo
Gaussian Wrapping
T 2 -Splat
Barn
0.51
0.67
0.49
0.47
0.52
0.67
0.47
Caterpillar
0.31
0.50
0.29
0.30
0.33
0.50
0.26
Courthouse
0.10
0.16
0.11
0.11
0.12
0.18
0.26
Ignatius
0.61
0.78
0.63
0.66
0.74
0.78
0.66
Meetingroom
0.19
0.41
0.24
0.25
0.23
0.41
0.21
Appendix
Table 8: Geometric reconstruction on Tanks & Temples. F-scores ( ↑ ) on six scenes and their arithmetic mean. The mean for T 2 -Splat is computed from unrounded per-scene scores.
Average faces (M)
Non-manifold edges (%) ↓
Non-manifold vertices (%) ↓
Triangle shape quality ↑
Faces in largest component (%)
A: Baseline
5.308
17.84
65.06
0.5796
52.94
B: Safe RDT
3.049
0
0
0.5918
44.30
C: Safe RDT + DTE
2.468
0
0
0.6203
54.41
Appendix
Table 9: Geometry and topology statistics of the final meshes on the nine scenes of Mip-NeRF 360. Triangle shape quality is reported as the equally weighted average of the per-scene median values.
Comparison
Δ PSNR (dB)
Δ LPIPS
Δ SSIM
Δ time (min)
(b) → (c): add DTE
+0.31
-0.02
+0.016
+14
(e) → (g): add DTE with 82 textures
+0.18
-0.01
+0.010
+2
(c) → (f): add 42 refinement
+0.31
-0.03
+0.019
+15
(f) → (g): increase 42 to 82
+0.06
-0.03
+0.004
+11
(a) → (g): complete configuration
+0.83
-0.07
+0.045
+29
Appendix
Table 10: Differences derived from the displayed values in Table 5 . Positive PSNR/SSIM changes and negative LPIPS changes indicate improvements. Time differences refer to reported total training time.
Truck
Train
PSNR ↑
22.91
19.32
LPIPS ↓
0.167
0.259
SSIM ↑
0.836
0.753
Appendix
Table 11: Per scene LPIPS, PSNR, and SSIM scores for the Truck and Train scenes of the T&T dataset.
Bicycle
Flowers
Garden
Stump
Treehill
PSNR ↑
23.96
19.81
25.68
25.36
21.01
LPIPS ↓
0.247
0.336
0.148
0.252
0.327
SSIM ↑
0.709
0.546
0.818
0.723
0.595
Appendix
Table 12: Per-scene PSNR, LPIPS, and SSIM scores for outdoor MipNeRF360 scenes.
Room
Counter
Kitchen
Bonsai
PSNR ↑
28.79
27.01
28.45
29.36
LPIPS ↓
0.224
0.214
0.150
0.214
SSIM ↑
0.904
0.885
0.909
0.924
Appendix
Table 13: Per-scene PSNR, LPIPS, and SSIM scores for indoor MipNeRF360 scenes.
Figure 4: Additional qualitative comparisons on Garden, Shelf, Truck, and Train. The columns show ground truth, T 2 -Splat, MeshSplatting, and MiLo. The highlighted regions provide close views of local texture and edge detail. Shelf is an additional STRinGS-360 example, outside the NVS averages reported in the per-scene tables.
Meshes remain the most practical representation for geometry reasoning and integration into graphics pipelines, yet existing reconstruction methods struggle to produce high-quality meshes. Most state-of-the-art approaches initially learn an intermediate representation (NeRF/3DGS) and treat mesh extraction as a post-processing step, which often leads to oversmoothed surfaces or poor quality meshes with excessive triangle counts.Existing mesh-native optimization methods alleviate some of these issues but suffer from fixed-resolution discretizations and unstable optimization behavior. In this paper, we introduce an adaptive mesh-based optimization framework and a practical mesh rendering technique to address these challenges. Our representation combines an optimizable Delaunay-triangulated tetrahedral grid with a multi-resolution hash grid. The former is refined through point pruning and insertion, while the latter provides latent features for SDF/appearance value predictions. We use volumetric rendering to bootstrap a coarse geometry while leveraging mesh-based rendering for recovering fine-grained details. Additionally, we propose a differentiable, rasterization-based depth-offset rendering formulation, reducing geometric artifacts and improving reconstruction quality. Our method significantly outperforms existing mesh optimization approaches across a variety of object-centric benchmarks while being competitive with state-of-the-art NeRF/3DGS methods.
Johannes Weidenfeller, Shaofei Wang, Philipp Fürnstahl +1
ETH Zurich, Zurich, Switzerland · Beijing Institute for General Artificial Intelligence, Beijing, China · University of Zurich, Zurich, Switzerland +1
Splatting-based radiance-field representations have recently gained popularity in robotic perception, enabling high-quality novel-view synthesis (NVS) through explicit scene primitives. However, existing splatting methods typically rely on iterative nonlinear optimization that jointly optimizes millions of coupled primitive parameters, resulting in substantial computation time, peak GPU memory consumption, and susceptibility to local minima. In this paper, we propose CF-Splat, a novel closed-form triangle-splatting framework for efficient scene appearance estimation over a known mesh either pre-existing or provided by upstream mapping or reconstruction pipelines. Our key insight is that, by properly formulating the problem over the given mesh geometry and calibrated cameras, the rendered pixel colors can be expressed linearly with respect to the per-vertex spherical harmonics (SH) coefficients, allowing multi-view scene appearance estimation to be formulated as a global linear least-squares (LLS) problem without the need of iterative nonlinear optimization. To guarantee a well-posed solution under uneven multi-view observation coverage, we introduce adaptive SH regularization that constrains insufficiently supported high-order SH coefficients and yields a unique global minimizer. To ensure the scalability to large-scale scenes, we develop a triangle-structured matrix-free preconditioned conjugate-gradient solver that avoids explicitly constructing the large sparse normal system while producing the same LLS solution. Experiments on diverse real-world datasets demonstrate that CF-Splat achieves competitive or superior NVS quality compared with state-of-the-art splatting methods. While, CF-Splat requires approximately 27.5-378.0 times less computation time and 1.8 times lower peak GPU memory on average than 3DGS across scenes of different scales on a consumer-grade laptop.
Triangle- and mesh-based neural rendering seeks to bridge neural representations and graphics engines through triangle primitives compatible with standard rasterization. Yet existing methods were developed and evaluated under inconsistent settings, with few direct comparisons and little evaluation in graphics engines. We introduce MeshSuite, the first benchmark for systematic evaluation of engine deployment and Non-Engine Render in triangle- and mesh-based neural rendering. MeshSuite evaluates two deployment settings: Standard Deployment, using an opaque mesh pipeline with vertex colors and hardware Z-buffering; and Dedicated Deployment, using method-specific engine implementations to preserve appearance and compositing features. For mesh splatting, we also audit exported surfaces for topological and geometric integrity as downstream graphics assets. Our results show that engine deployment reduces image quality in both settings; among evaluated methods, MeshSplatting degrades least under Standard Deployment. Dedicated Deployment preserves most fidelity at roughly 5-33x the runtime of Standard Deployment across two deployment datasets. For mesh splatting, explicit connectivity and shared indexing remain limited: shared vertex indexing alone guarantees neither manifoldness nor global connectivity. These findings show that rasterizability is only part of graphics readiness and underscore the need to assess engine deployment. Project page: https://x-up-lab.github.io/research/meshsuite.
Kaixuan Zhang, Minxian Li, Mingwu Ren +1
Nanjing University of Science and Technology · University of Surrey