3D Gaussian Splatting has made Gaussian primitives a highly efficient representation for real-time novel view synthesis, but its rasterisation-based formulation relies on screen-space approximations that limit accurate view-dependent ordering and the integration of secondary ray effects such as reflections, refractions, and shadows. Gaussian ray tracing addresses these limitations by evaluating explicit ray-primitive intersections, yet it remains costly to train. We observe that the main bottleneck is not ray traversal alone, but the pixel-centric backward propagation, where many threads concurrently accumulate gradients into the same primitive parameters, causing severe atomic contention and thread serialisation. We present 3DGART, a practical training framework for ray-traced Gaussian rendering. Our key idea is to reorganise backward propagation around primitives rather than pixels. Using conservative perspective-correct screen-space bounds, we build a compact intermediate buffer and a tile-primitive mapping that allows each thread to accumulate the contribution of one primitive over its covered pixels within a tile. This transforms gradient computation from a contention-heavy scatter operation into a structured gather-like process. On Mip-NeRF 360, 3DGART achieves an ≈3−3.5× raw training speedup over per-pixel baseline and ≈4× over 3DGRT on Mip-NeRF 360 while improving quality. More importantly, 3DGART makes fully ray-traced Gaussian training practical, reaching runtimes competitive with rasterisation-based pipelines while preserving benefits of ray tracing.
Figures & tables
Figure 1. NVIDIA 3DGRT improves efficiency through compromises, including a shorter Gaussian kernel, a stricter opacity cutoff, and capped ray-primitive intersections. In contrast, 3DGART (Ours) preserves the ray-traced Gaussian formulation and accelerates training by reorganizing the backward pass. On the Mip-NeRF 360 bicycle scene, 3DGART converges ≈4× faster than 3DGRT, while achieving higher reconstruction quality.
Figure 2. Forward pass. Each ray–primitive intersection stores the accumulated colour and transmittance tuple L,T) in primitive-major tile-local order within our intermediate buffer.
Figure 3. Backward pass: tile-particle mapping. Each thread iterates over a strided subset of primitives from the tile primitive list. For each primitive, it retrieves the corresponding LT buffer offset, accumulates gradients over the covered pixels, and writes the result to global memory using atomicAdd.
Mip-NeRF 360
#G
Pixel
Hybrid
Primitive
Gain ↑
1.00M
44m08s
19m55s
12m35s
3.51 ×
2.00M
51m13s
24m56s
16m32s
3.10 ×
Table 1. Raw head-to-head training-speed benchmark on Mip-NeRF 360. Gain is measured against the corresponding pixel-based representation.
Mip-NeRF 360
Tanks&Temples
Deep Blending
Model
Type
Iter
PSNR ↑
SSIM ↑
LPIPS ↓
Train
#G
PSNR ↑
SSIM ↑
LPIPS ↓
Train
#G
PSNR ↑
SSIM ↑
LPIPS ↓
Train
#G
ZipNeRF
–
–
28.54
0.828
0.219
5h
–
–
–
–
–
–
–
–
–
–
–
3DGS
Rast.
30K
27.35
0.814
0.216
20m19s
3.31M
23.52
0.842
0.183
11m53s
1.84M
29.77
0.906
0.263
20m55s
2.81M
3DGUT
Rast.
30K
27.47
0.812
0.218
16m09s
3.23M
23.52
0.848
0.173
11m24s
2.15M
29.60
0.902
0.256
8m35s
1.37M
HTGS
Rast.
30K
27.04
0.818
0.196
10m09s
2.20M
23.09
0.846
0.172
6m32s
0.77M
29.93
0.907
0.234
8m25s
1.01M
3DGRT
RT
30K
27.11
0.809
0.214
62m21s
4.19M
22.81
0.844
0.165
36m47s
4.05M
29.62
0.904
0.247
50m51s
1.74M
Table 2. Global benchmark across Mip-NeRF 360, Tanks&Temples, and Deep Blending. Our approach significantly reduces training time over the current state-of-the-art fully ray tracing approach (3DGRT), while matching rasterisation-based methods.
Figure 4. Per-component iteration time (ms). Runtime breakdown for the per-pixel baseline, our hybrid approach, and our tile-primitive method at 1.0×106 and 2.0×106 primitives, highlighting the reduction in backward pass overhead.
Figure 5. Peak memory analysis. Comparison of total VRAM usage and LT buffer consumption across primitive scales ( 1M to 5M ) for our tile–primitive approach at low and high resolutions (up to 1.5 MP ). For the hybrid approach, the low and high settings correspond to resolutions up to 1.5 MP and 6 MP , respectively.
3D Gaussian Splatting (3DGS) achieves real-time novel-view synthesis by optimizing millions of anisotropic Gaussians, yet its training remains expensive, with the backward pass dominating runtime in the post-densification refinement phase. We observe substantial update redundancy in this phase: many sampled views have near-plateaued losses and provide diminishing gradient benefits, but standard training still runs full backpropagation. We propose SkipGS with a novel view-adaptive backward gating mechanism for efficient post-densification training. SkipGS always performs the forward pass to update per-view loss statistics, and selectively skips backward passes when the sampled view's loss is consistent with its recent per-view baseline, while enforcing a minimum backward budget for stable optimization. On Mip-NeRF 360, compared to 3DGS, SkipGS reduces end-to-end training time by 23.1%, driven by a 42.0% reduction in post-densification time, with comparable reconstruction quality. Because it only changes when to backpropagate without modifying the renderer, representation, or loss, SkipGS is plug-and-play and compatible with other complementary efficiency strategies, enabling additive speedups. Code is available at https://github.com/ASU-ESIC-FAN-Lab/SkipGS.
Consumer-level applications require fast optimization of 3D Gaussian Splatting (3DGS) with high-fidelity novel view rendering. However, existing 3DGS acceleration approaches still incur substantial computation on redundant pixels while sacrificing fine details. In this paper, we present TurboGS, an error-guided training framework that accelerates 3DGS by concentrating optimization on perceptually informative pixels. TurboGS is built upon four core components: (1) a tile-wise sparse pixel sampling, which, driven by multi-view reconstruction errors during training, prioritizes challenging regions and skips well-reconstructed ones to avoid redundant gradient computation; (2) a tile-wise structure-aware loss with sparse Normalized Cross-Correlation, which provides sparse yet effective supervision to preserve fine details and stabilize training; (3) an error-driven Gaussian density control strategy, which dynamically allocates model capacity and removes redundant primitives; and (4) a tailored hybrid optimizer that couples Hessian-informed updates with Adam moment damping to stabilize and improve convergence under sparse supervision. Experiments on standard benchmarks demonstrate that TurboGS can deliver on par or superior rendering quality within 100 seconds on a single RTX 5090 GPU card (up to 10x training speedup over vanilla 3DGS).
Zheng Dong, Daifei Qiu, Pinxuan Dai +5
Zhejiang Sci-Tech University · Malanshan Audio&Video Laboratory · CAD&CG Laboratory, Zhejiang University +2
Training 3D Gaussian Splatting (3DGS) at billion-primitive scale is fundamentally memory-bound: each Gaussian primitive carries a large attribute vector, and the aggregate parameter table quickly exceeds GPU capacity, limiting prior systems to tens of millions of Gaussians on commodity single-GPU hardware. We observe that 3DGS training is inherently sparse and trajectory-conditioned: each iteration activates only the Gaussians visible from the current camera batch, so GPU memory can serve as a working-set cache rather than a persistent parameter store. Building on this insight, we introduce TideGS, an out-of-core training framework that manages parameters across an SSD-CPU-GPU hierarchy via three synergistic techniques: block-virtualized geometry for SSD-aligned spatial locality, a hierarchical asynchronous pipeline to overlap I/O with computation, and trajectory-adaptive differential streaming that transfers only incremental working-set deltas between iterations. Experiments show that TideGS enables training with over one billion Gaussians on a single 24 GB GPU while achieving the best reconstruction quality among evaluated single-GPU baselines on large-scale scenes, scaling beyond prior out-of-core baselines (e.g., approximately 100M Gaussians) and standard in-memory training (e.g., approximately 11M Gaussians).
Chonghao Zhong, Linfeng Shi, Hua Chen +4
Hong Kong University of Science and Technology · Great Wall Motor · Tsinghua University +1