AdaptSplat: Adaptive 3D Gaussian Splatting

• AdaptSplat is a feed-forward 3D Gaussian splatting method that adapts a vision foundation model using a lightweight Frequency-Preserving Adapter to restore high-frequency detail.
• It maintains the standard pipeline of image feature extraction, multi-view interaction, and feature decoding while injecting high-frequency priors into both attention and decoding to counteract low-pass filtering.
• The approach demonstrates improved reconstruction metrics and cross-domain generalization on multiple benchmarks through end-to-end backbone adaptation and DWT-derived priors.

Searching arXiv for AdaptSplat and closely related feed-forward 3D Gaussian splatting work. AdaptSplat is a feed-forward 3D Gaussian Splatting method that adapts a vision foundation model through a single lightweight adapter rather than through complex, architecture-specific redesign. It retains the generic pipeline of image feature extraction, multi-view interaction, and feature decoding, but augments it with a 1.5 M-parameter Frequency-Preserving Adapter (FPA) that extracts direction-aware high-frequency structural priors from shallow backbone features and reinjects them into both attention and decoding. In the reported formulation, the method uses a fully unfrozen DINOv3-ConvNeXt backbone, a standard multi-view Transformer, a DPT decoder, and light Gaussian-parameter prediction heads, and is presented as addressing both cross-domain generalization and the loss of high-frequency geometric detail caused by deep-network low-pass filtering (Xing et al., 11 May 2026).

1. Position within feed-forward 3D Gaussian Splatting

AdaptSplat is framed around a now-standard feed-forward 3DGS pipeline: image feature extraction $\rightarrow$ multi-view interaction $\rightarrow$ feature decoding. The central claim is that this generic pipeline does not require extensive 3D-specific component engineering in order to improve reconstruction quality and zero-shot generalization. Instead, the method inserts a single adapter family into the backbone and uses its outputs at two later stages of the pipeline (Xing et al., 11 May 2026).

The problem setting is specifically feed-forward 3D Gaussian Splatting rather than optimization-based 3DGS. AdaptSplat predicts Gaussian primitives directly from multi-view image features. The prediction heads regress each Gaussian’s position $\mu$, opacity $\alpha$, color $\mathbf{c}$, scale $\mathbf{s}$, and rotation quaternion $\mathbf{q}$, after which a differentiable rasterizer renders novel views. The method attributes two recurrent deficiencies in prior feed-forward systems to the underlying feature hierarchy: first, deep networks smooth away high-frequency edges, producing “safe” but isotropic Gaussians that blur sharp boundaries; second, freezing a foundation-model backbone limits the adaptation of its priors to 3D geometry, which in turn constrains cross-domain generalization (Xing et al., 11 May 2026).

This design places AdaptSplat in a specific methodological niche. It is neither a new multi-view interaction module nor a new decoder family. A plausible implication is that the work treats feature adaptation, rather than architectural replacement, as the main remaining leverage point in feed-forward 3DGS.

2. Architectural pipeline and insertion points

The base architecture consists of four stages. A DINOv3-ConvNeXt backbone, unfrozen for end-to-end training, extracts multi-scale 2D features. A standard multi-view Transformer aggregates these features across viewpoints. A DPT decoder upsamples and refines the aggregated features. Finally, light prediction heads output Gaussian parameters for differentiable rendering (Xing et al., 11 May 2026).

AdaptSplat modifies this architecture at two locations. First, it inserts four parallel Frequency-Preserving Adapter modules, one at each resolution of the ConvNeXt backbone. Second, it uses the resulting high-frequency priors in two distinct injection mechanisms. In the multi-view Transformer, the priors are treated as positional encodings in query/key space, with the stated effect of sharpening attention around edges. In the DPT decoder, the same priors generate spatial gating masks for adaptive residual modulation, which is intended to counteract the low-pass effect of bilinear upsampling (Xing et al., 11 May 2026).

Implementation details further specify the scale of the intervention. The backbone is DINOv3-ConvNeXt and is fully unfrozen. The Transformer is standard multi-view “Vggt” attention with $d=512$ feature dimension. The decoder is DPT with four upsampling levels. The adapter parameters are distributed roughly as 0.1 M, 0.3 M, 0.5 M, and 0.6 M across the four FPA sub-modules, following ConvNeXt’s channel pyramid $96 \rightarrow 192 \rightarrow 384 \rightarrow 768$ (Xing et al., 11 May 2026).

A common misconception is to view AdaptSplat as a frozen-backbone adapter method. The reported implementation is explicitly end-to-end with a fully unfrozen backbone, and the adapter is presented as a lightweight complement to full backbone adaptation rather than as a substitute for it.

3. Frequency-Preserving Adapter

The Frequency-Preserving Adapter operates on shallow ConvNeXt feature maps $$F_{\mathrm{shallow}} \in \mathbb{R}^{C\times H\times W}$$ and has two components: direction-aware high-frequency prior extraction via a 2D discrete wavelet transform, and a lightweight convolutional bottleneck that projects the resulting priors back to the original channel dimension (Xing et al., 11 May 2026).

For high-frequency prior extraction, the method applies orthogonal low-pass and high-pass 1D wavelet filters $\rightarrow$0 and $\rightarrow$1 to form the four standard 2D subbands:

$\rightarrow$2

The informative components are $\rightarrow$3, $\rightarrow$4, and $\rightarrow$5, interpreted respectively as horizontal edges, vertical edges, and diagonal textures. These are concatenated along the channel dimension:

$\rightarrow$6

The projection-and-refinement step upsamples the wavelet tensor and passes it through a two-layer bottleneck with reduction ratio $\rightarrow$7:

$\rightarrow$8

Here, $\rightarrow$9 denotes nearest-neighbor upsampling by a factor of 2, $\mu$0 and $\mu$1 reduce channels, $\mu$2 is ReLU, and $\mu$3 with $\mu$4 restore the original channel count. The output $\mu$5 is the refined high-frequency prior (Xing et al., 11 May 2026).

The parameterization is deliberately small. For one sub-module, the count is

• $\mu$6,
• $\mu$7.

Summed over the four ConvNeXt stages with $\mu$8, this yields approximately 1.5 M total adapter parameters. The design objective is not merely compactness. The paper argues that these priors preserve directional high-frequency information that is otherwise attenuated by the spectral bias of deep features (Xing et al., 11 May 2026).

4. Injection into attention and decoding

AdaptSplat uses the FPA output in two different ways. In the multi-view Transformer, $\mu$9 is treated as a high-frequency positional encoding rather than concatenated directly to the backbone features. After normalization and an optional fixed sinusoidal mapping,

$\alpha$0

the prior is added channel-wise to $\alpha$1 and $\alpha$2 in self-attention. The stated rationale is that this preserves the semantic subspace of deep features while biasing attention similarity toward edge-aligned regions (Xing et al., 11 May 2026).

In the DPT decoder, the same prior generates spatial gating masks

$\alpha$3

where $\alpha$4 is the sigmoid. If $\alpha$5 denotes the upsampled deep feature and $\alpha$6 the shallower feature at scale $\alpha$7, the decoder uses adaptive residual modulation:

$\alpha$8

with $\alpha$9 a learned scalar per scale. The intended effect is to amplify shallow residual features where the high-frequency prior is strong, thereby counteracting blur introduced by bilinear upsampling (Xing et al., 11 May 2026).

The optimization setup combines a reconstruction term, a frequency-domain term, and a regularizer:

$\mathbf{c}$0

The reconstruction loss is

$\mathbf{c}$1

the frequency-focused loss is

$\mathbf{c}$2

and $\mathbf{c}$3 is a small opacity regularizer. The coefficients are reported as $\mathbf{c}$4 with $\mathbf{c}$5. Optimization uses AdamW with weight decay $\mathbf{c}$6, backbone learning rate $\mathbf{c}$7 and the remainder at $\mathbf{c}$8 in Stage 1, then uniform $\mathbf{c}$9 thereafter, under cosine annealing with 3 k warmup steps (Xing et al., 11 May 2026).

5. Training regime, benchmarks, and ablations

The reported training schedule spans multiple datasets. On DL3DV, training proceeds in three stages at resolutions $\mathbf{s}$0 for 100 k iterations, $\mathbf{s}$1 for 50 k iterations, and $\mathbf{s}$2 with dynamic view counts for 30 k iterations. On RealEstate10K, training uses $\mathbf{s}$3 images with 6 input and 8 target views and the same hyperparameters as DL3DV Stage 1. Zero-shot inference is reported on Tanks & Temples and MipNeRF360 without fine-tuning (Xing et al., 11 May 2026).

The main quantitative results are as follows:

Setting	Representative prior result	AdaptSplat result
RealEstate10K, 6 in $\mathbf{s}$4 8 novel, $\mathbf{s}$5	MVP: 32.89 / 0.948 / 0.067	AdaptSplat (tiny): 33.70 / 0.955 / 0.063; AdaptSplat (base): 33.86 / 0.956 / 0.062
DL3DV, 32 views	MVP: 25.96 / 0.847 / 0.187	AdaptSplat: 26.28 / 0.860 / 0.177
DL3DV, 128 views	MVP: 29.02 / 0.903 / 0.134	AdaptSplat: 29.27 / 0.911 / 0.127
Tanks & Temples, zero-shot, 128 views	MVP: 22.36 / 0.804 / 0.184	AdaptSplat: 22.81 / 0.826 / 0.171
MipNeRF360, zero-shot, 128 views	MVP: 25.12 / 0.736 / 0.248	AdaptSplat: 25.60 / 0.756 / 0.231

These metrics are reported as PSNR / SSIM / LPIPS, with higher PSNR and SSIM and lower LPIPS preferred (Xing et al., 11 May 2026).

Ablation studies isolate the contribution of each design choice. On a 2 k-scene DL3DV subset trained for 50 k iterations, the progression is: baseline with frozen ConvNeXt, PSNR 21.12; plus unfrozen DINOv3, 21.47; plus FPA-guided attention positional encoding, 21.75; plus FFL loss, 21.83; plus decoder residual modulation, 22.10. For high-frequency prior extraction, DWT is compared against Fourier filtering, learned convolutional high-pass filters, and Sobel edges; the DWT variant reports 21.75/0.268/0.711 against 18.37/0.477/0.529, 18.39/0.481/0.535, and 21.45/0.289/0.609, respectively. Transformer fusion by positional-encoding injection is compared with simple additive fusion: Add gives PSNR 21.16 / LPIPS 0.287 / SSIM 0.687, whereas PE gives 21.47 / 0.291 / 0.699 (Xing et al., 11 May 2026).

The paper also reports a change in the fractional anisotropy of Gaussian covariance on RealEstate10K: 0.8015 without FPA and 0.8423 with FPA, described as indicating more anisotropic, edge-aligned Gaussians. Efficiency on RealEstate10K with 6 views is reported as 241 M parameters and 0.039 s/frame for MVP at PSNR 32.89, 215 M and 0.042 s/frame for AdaptSplat (tiny) at PSNR 33.70, and 420 M and 0.061 s/frame for AdaptSplat (base) at PSNR 33.86 (Xing et al., 11 May 2026).

6. Nomenclature, related methods, and limitations

Despite the similarity in naming, AdaptSplat is distinct from several other “Splat”-named systems. “MuSASplat: Efficient Sparse-View 3D Gaussian Splats via Lightweight Multi-Scale Adaptation” freezes a ViT encoder, injects a Multi-Scale Adapter into each Transformer block, and uses a one-shot Feature Fusion Aggregator for pose-free sparse-view feed-forward 3DGS (Xu et al., 8 Dec 2025). “AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting” instead modifies adaptive density control in vanilla 3DGS through an error-driven split operator that determines the number of split children from pixel-error region statistics (Lee et al., 7 May 2026). “AdaSplats: Adaptive Splatting of Point Clouds for Accurate 3D Modeling and Real-time High-Fidelity LiDAR Simulation” concerns adaptive splat generation from point clouds for LiDAR simulation rather than feed-forward image-based Gaussian prediction (Richa et al., 2022). “SPLAT: Semantic Pixel-Level Adaptation Transforms for Detection” is an unsupervised domain-adaptation method for object detection and is unrelated to 3D Gaussian Splatting despite the acronym overlap (Tzeng et al., 2018).

This naming overlap can obscure the specific contribution of AdaptSplat. Its distinctive claim is not adaptive splitting, point-cloud splat generation, or pixel-level image translation, but adaptation of a vision foundation model for feed-forward 3DGS via a frequency-preserving adapter. Within the feed-forward 3DGS literature represented here, a useful contrast is that MuSASplat emphasizes lightweight fine-tuning of a frozen ViT and efficient multi-view fusion under sparse views, whereas AdaptSplat emphasizes end-to-end backbone adaptation and explicit restoration of high-frequency detail through DWT-derived priors (Xu et al., 8 Dec 2025).

The limitations reported for AdaptSplat are specific. The method relies on fixed DWT filters, and the discussion notes that future work might learn hybrid wavelets or trainable frequency bases. It also assumes sufficiently detailed shallow feature maps; extending to extremely sparse views or lower-resolution backbones may require more advanced multi-scale fusion. Suggested extensions include learned wavelet banks that adapt to scene statistics, cross-modal adapters that fuse depth or surface-normal priors, and sparse or point-based injection for lighter-weight inference (Xing et al., 11 May 2026).

Taken together, the reported results position AdaptSplat as a minimal adapter-based formulation of feed-forward 3DGS: a method in which frequency-preserving adaptation, rather than wholesale pipeline reinvention, is presented as sufficient to improve both reconstruction fidelity and cross-domain generalization (Xing et al., 11 May 2026).