Pose-free feed-forward 3D Gaussian Splatting enables novel view synthesis from uncalibrated multi-view images. Although more views should improve performance, existing methods often degrade with dense-view inputs because global aggregation spreads attention over many tokens, and naive voxel fusion averages many Gaussians into overly smooth representations. We present AVSplat, a framework that turns additional views into reliable signals for both aggregation and representation. Before global attention, each view performs a single lightweight interaction with a small set of Assist Views chosen for relevance and diversity, and the cached features provide a focused scene context that stabilizes correspondence. For representation, we use adaptive temperature-aware voxel fusion that sharpens attribution under high occupancy, guided by occupancy and point confidence. Crucially, AVSplat restores positive view scaling where performance remains stable or improves as more input views are added, instead of degrading in the dense-view regime. Ablations show that Assist View Preconditioning is primarily responsible for preventing dense-view degradation, while Occupancy-guided Voxel Fusion contributes most of the single-point image-quality gains.
Figures & tables
Figure 1: Overview of the AVSplat pipeline. The encoder converts each input image into patch embeddings (image tokens) and aggregates them into per-view tokens. Assist-View Preconditioning (AVP) computes a descriptor for each view and injects scene context before global aggregation. Decoder heads then predict camera parameters, depth, and Gaussian primitives. Occupancy-guided Voxel Fusion (OVF) forms voxel-level attributes and modulates them by voxel occupancy, followed by differentiable Gaussian rasterization to produce the final rendering.
Sparse inputs
3
6
16
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Mip-NeRF 360 [ 1 ]
NoPoSplat [ 22 ]
16.36
0.430
0.453
15.92
0.416
0.541
15.47
0.361
0.606
Flare [ 26 ]
13.52
0.350
0.601
15.35
0.407
0.573
13.21
0.348
0.695
AnySplat [ 5 ]
17.29
0.429
0.433
17.51
0.446
0.412
17.81
0.448
0.382
Table 1: Quantitative results on Mip-NeRF 360 [ 1 ] , VR-NeRF [ 20 ] , and DL3DV [ 9 ] under sparse and dense inputs with metrics PSNR ↑ , SSIM ↑ [ 17 ] , LPIPS ↓ [ 25 ] . The best result in each group is highlighted in red and the second best in orange. AVSplat attains rendering quality comparable to the strongest baselines and avoids the dense-view degradation observed in prior pose-free pipelines. Our performance remains stable or improves as more input views are added, instead of oscillating or degrading at high view counts.
Figure 2: Qualitative comparison between AVSplat and AnySplat [ 5 ] under dense-view inputs. We focus on the dense-view setting, where the artifacts we target primarily occur, and therefore omit sparse-view results. Red boxes highlight that our method preserves more fine-grained surface details with fewer artifacts and less oversmoothing-induced blur, while the rightmost column shows more accurate global matching.
Figure 3: Cross-view attention at layer 15 for a fixed frame pair under 16-view inputs. The maps are extracted from the full attention matrix after removing self-frame blocks and special tokens, and are displayed with identical normalization. AVSplat produces more concentrated cross-view responses and the geometric consistency is quantified in Tab. 3 .
Figure 4: Performance versus input count with 3, 6, 16, 32, 48, 64, and 72 input views. The vertical axis reports PSNR and the horizontal axis shows the number of input views. Shaded regions indicate the standard deviation across multiple trials. In the dense-view setting, AVSplat consistently outperforms AnySplat [ 5 ] . The inset in the lower right zooms into the dense-view region and shows that AnySplat begins to oscillate as the number of input views increases, whereas AVSplat maintains stable or improving performance without degradation.
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
Baseline
21.14
0.684
0.282
+ Fusion only
21.57
0.692
0.272
+ Assist + Fusion
21.67
0.717
0.265
Table 2: Ablations on the DL3DV [ 9 ] in the 64 views setting. Occupancy-guided Voxel Fusion contributes most of the single-point image-quality improvement at 64 views, while Assist-View Preconditioning mainly targets the dense-view scaling failure mode by stabilizing global matching as the input count grows.
Attention diagnostics
ACID overlap stress test
Method
Entropy ↓
GeoMass ↑
Method
48 views
72 views
AnySplat
0.693
0.286
AnySplat
23.12
22.45
AVSplat
0.642
0.374
AVSplat
23.15
23.24
Table 3: Additional dense-view diagnostics. Left: cross-view attention analysis on DL3DV with 16 input views. GeoMass@1patch measures the attention assigned to a one-patch-width epipolar band. Right: PSNR (dB) on the overlap-heavy ACID evaluation. Ground-truth poses are used only to compute the diagnostic metric and are never provided to either model.
Department of Electrical and Computer Engineering (ECE), Sungkykwan University (SKKU), Suwon 16419, South Korea · Department of Display Convergence Engineering (DCE), SKKU, Suwon 16419, South Korea · AiM Future, Seoul 06804, South Korea +1