Feed-forward 3D Gaussian Splatting enables efficient generalizable novel view synthesis, but existing pixel-aligned methods often generate highly redundant Gaussian primitives, causing the final representation size and rendering cost to grow rapidly with image resolution and the number of input views. We propose BudgetSplat, a budget-controllable feed-forward framework that adapts the final Gaussian representation to a user-specified preservation ratio. Rather than treating compaction as simple importance ranking, we formulate it as a joint problem of importance estimation, spatial coverage, and cross-view redundancy. A single trained model supports varying numbers of context views and multiple test-time budgets without per-budget retraining or per-scene compression. To realize this formulation, Intra-Cross Mask Supervision provides structure-, coverage-, and redundancy-aware guidance for learning Gaussian importance, while budget-conditioned feature modulation adapts Gaussian opacity and attributes to the requested preservation ratio. Coverage-Aware Budget Selection further combines patch-wise pre-selection, view-wise opacity ranking, and 3D redundancy suppression to construct compact yet spatially distributed Gaussian representations. Experiments on indoor reconstruction and cross-dataset generalization demonstrate consistent quality--size trade-offs across different input view counts and preservation ratios, substantially reducing final representation storage and rendering cost while preserving rendering quality.
Figures & tables
Figure 1: Overview of BudgetSplat . (A) Given N posed context images, a pixel-aligned reconstruction backbone estimates dense geometry and base Gaussian features. (B) During training, Intra-Cross Mask Supervision constructs pseudo importance targets that preserve high-frequency structures, maintain sparse coverage in low-frequency regions, and down-weight redundant cross-view observations. (C) Budget-Conditioned Feature Modulation injects the user-specified preservation ratio ρ into the base Gaussian features before parameter prediction, allowing the predicted opacity and Gaussian attributes to adapt to the requested budget. (D) Coverage-Aware Budget Selection combines patch-wise pre-selection, view-wise global opacity ranking, and 3D redundancy suppression to retain a compact and spatially distributed Gaussian representation. A single trained model supports arbitrary test-time preservation ratios ρ∈(0,1] without retraining.
Figure 2: Qualitative results on ScanNet [ 6 ] with 10 context views. Ours is rendered using only 25% of the Gaussian budget.
Type
Method
2 Views
3 Views
PSNR ↑
SSIM ↑
LPIPS ↓
#GS(k) ↓
PSNR ↑
SSIM ↑
LPIPS ↓
#GS(k) ↓
Native capacity
pixelSplat [ 2 ]
24.50
0.776
0.288
1180
23.44
0.782
0.270
1769
MVSplat [ 5 ]
24.02
0.771
0.268
393
22.26
0.744
0.328
590
DepthSplat [ 23 ]
24.59
0.783
0.276
393
23.10
0.761
0.327
590
FreeSplat [ 20 ]
28.08
0.837
0.211
278
27.45
0.829
0.222
382
FreeSplat++ † [ 21 ]
28.19
0.838
0.212
68
27.45
0.829
0.223
89
Table 1: Generalizable NVS interpolation results on ScanNet [ 6 ] with 2 and 3 context views. The input resolution is 384×512 . † denotes results reported in the original paper.
Type
Method
#GS(k) ↓
GS Size (MiB) ↓
View Interpolation
View Extrapolation
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Native capacity
pixelSplat [ 2 ]
1966
442.5
20.31
0.713
0.375
18.93
0.680
0.410
MVSplat [ 5 ]
655
147.4
19.25
0.661
0.379
19.00
0.663
0.382
DepthSplat [ 23 ]
655
147.4
21.76
0.735
0.297
21.10
0.720
0.311
FreeSplat [ 20 ]
294
66.2
25.03
0.800
0.223
23.96
0.780
0.250
ZPressor [ 18 ]
393
88.5
21.59
0.746
0.317
21.43
0.730
0.289
Table 2: Generalizable NVS interpolation and extrapolation results on ScanNet [ 6 ] with 10 context views. The input resolution is 256×256 .
Type
Method
3 Views
10 Views
PSNR ↑
SSIM ↑
LPIPS ↓
#GS(k) ↓
PSNR ↑
SSIM ↑
LPIPS ↓
#GS(k) ↓
Native capacity
pixelSplat [ 2 ]
25.88
0.838
0.214
1769
21.03
0.745
0.344
1966
MVSplat [ 5 ]
23.81
0.812
0.221
590
19.11
0.696
0.348
655
DepthSplat [ 23 ]
25.73
0.844
0.169
590
20.29
0.733
0.275
655
FreeSplat [ 20 ]
26.98
0.848
0.171
423
21.24
0.752
0.289
448
ZPressor [ 18 ]
23.93
0.816
0.197
590
21.23
0.750
0.258
393
Table 3: Cross-dataset generalization on Replica [ 16 ] . The input resolution is 384×512 for the 3-view setting and 256×256 for the 10-view setting.
Method
Components
PSNR ↑
SSIM ↑
LPIPS ↓
#GS(k) ↓
Baseline
13.02
0.386
0.561
147
No.1
I
24.80
0.743
0.336
147
No.2
I + A
26.14
0.810
0.311
147
No.3
I + C + A
26.38
0.813
0.310
147
No.4
I + A + S
26.75
0.821
0.279
147
Full
I + C + A + S
26.77
0.823
0.275
147
Table 4: Ablation on ScanNet [ 6 ] . I, C, A, and S denote intra-frame supervision, cross-frame weighting, alpha completion, and coverage-aware budget selection, respectively. Results are reported using 3 context views.