Frequency-Preserving Adapter (FPA)

• Frequency-Preserving Adapter (FPA) is a lightweight module that recovers high-frequency geometric details in feed-forward 3D Gaussian Splatting via wavelet decomposition.
• It extracts direction-aware priors from shallow DINOv3-ConvNeXt features and injects them into multi-view attention and decoder modulation for improved feature preservation.
• Empirical results show that FPA boosts metrics like PSNR, LPIPS, and SSIM while reducing isotropic degeneration to maintain sharp geometric boundaries.

Searching arXiv for the cited FPA paper and closely related frequency-aware adapter work to ground the article in current literature. Frequency-Preserving Adapter (FPA) is a lightweight adapter design introduced in AdaptSplat for feed-forward 3D Gaussian Splatting, where it serves to recover and inject high-frequency structural information that is attenuated by the low-pass filtering effect of deep networks (Xing et al., 11 May 2026). In that formulation, FPA extracts direction-aware high-frequency structural priors from shallow DINOv3-ConvNeXt features and injects them into a generic pipeline of image feature extraction, multi-view interaction, and feature decoding through high-frequency positional encodings and adaptive residual modulation. The term denotes preservation of geometry-relevant high-frequency structure rather than exact conservation of the full spectral content, and it should therefore be distinguished from neighboring frequency-aware adapter designs that emphasize routing, activation, or feature injection rather than explicit “frequency-preserving” behavior (Xing et al., 11 May 2026).

1. Terminology, scope, and disambiguation

In current adapter literature, “Frequency-Preserving Adapter” refers most concretely to the module in AdaptSplat, where the adapter is designed to compensate for high-frequency attenuation in feed-forward 3DGS and thereby improve high-frequency geometric fidelity (Xing et al., 11 May 2026). The preservation target is not an invariant spectral identity map. Rather, the adapter preserves access to local boundaries, thin structures, and direction-aware high-frequency cues that would otherwise be suppressed as features traverse a deep backbone, a multi-view Transformer, and a decoder.

This usage is narrower than several adjacent notions in the literature. FeRA is described as a frequency-aware, frequency-aligned LoRA expert routing framework for diffusion fine-tuning rather than a literal preservation method (Yin et al., 22 Nov 2025). FAA is a Fourier-Activated Adapter that uses random Fourier features, adaptive weighting, and gating for LLM post-training, but it is described as frequency-aware activation and fusion rather than strict spectral preservation (Bae et al., 26 Dec 2025). FSAM introduces a frequency adapter for medical image segmentation that injects Fourier-amplitude information into SAM, again as aggregation and adaptation rather than exact preservation (Bui et al., 11 May 2026). Earth-Adapter uses a mixture of frequency adaptation for remote-sensing segmentation, and VMDA uses a frequency-guided visual adapter for multi-modal tracking; both are relevant neighboring designs rather than the same formulation (Hu et al., 8 Apr 2025, Xu et al., 30 Jun 2025).

The acronym is also ambiguous. In “A Light-Weight Object Detection Framework with FPA Module for Optical Remote Sensing Imagery,” FPA does not denote Frequency-Preserving Adapter at all; it refers in practice to an FPN-plus-attention feature aggregation module inside CenterFPANet (Gu et al., 2020). Any encyclopedia treatment of FPA therefore requires explicit contextual disambiguation.

2. Problem setting in feed-forward 3D Gaussian Splatting

AdaptSplat motivates FPA by identifying a specific failure mode in the generic feed-forward 3DGS pipeline,

$$\text{image feature extraction} \rightarrow \text{multi-view interaction} \rightarrow \text{feature decoding},$$

instantiated with DINOv3-ConvNeXt for feature extraction, a standard multi-view Transformer for interaction, and a standard DPT decoder for feature decoding (Xing et al., 11 May 2026). The paper argues that deep networks exhibit a low-pass filtering effect, so local sharp details are attenuated as processing moves toward deeper semantic features. In 3D reconstruction, this attenuation degrades the evidence available for local boundary orientation and fine geometric structure.

The paper ties this directly to Gaussian geometry. In 3DGS, the Gaussian covariance is

$$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$

and the projected 2D covariance is

$$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$

When boundary direction is uncertain because the feature hierarchy has become over-smoothed, the model tends to regress

$s_x \approx s_y \approx s_z,$

producing an isotropic Gaussian sphere rather than an anisotropic primitive aligned with a surface or contour (Xing et al., 11 May 2026). The paper terms this isotropic degeneration and presents it as a principal reason for blurred geometry near edges, thin structures, and sharp boundaries.

FPA is introduced as the only major custom module added to an otherwise generic feed-forward 3DGS architecture. Its purpose is to break this degeneration by reintroducing directional high-frequency priors derived from shallow backbone features, which the paper argues retain local structure that deep features lose (Xing et al., 11 May 2026).

3. Architectural formulation

FPA is a single lightweight adapter of about $1.5$M parameters (Xing et al., 11 May 2026). It takes shallow features from DINOv3-ConvNeXt and extracts high-frequency priors $\mathbf{F}_{hf}$, which are then consumed in two places: the multi-view Transformer and the DPT decoder. The paper states that FPA “directly extracts direction-aware high-frequency structural priors from shallow features” and “introduces 2D DWT to break this degeneration” (Xing et al., 11 May 2026).

The wavelet decomposition is central to the design. FPA uses 2D DWT to produce four subbands:

• $\mathbf{LL}$: low-low,
• $\mathbf{LH}$: horizontal,
• $\mathbf{HL}$: vertical,
• $\mathbf{HH}$: diagonal.

The high-frequency bands $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$0, $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$1, and $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$2 carry orientation-sensitive local structure. The paper specifically states that $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$3 captures horizontal high-frequency content, $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$4 captures vertical high-frequency content, and $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$5 captures diagonal high-frequency content (Xing et al., 11 May 2026). This is the basis for the descriptor “direction-aware”: the prior is not merely high-frequency energy, but orientation-sensitive structural evidence.

Rather than concatenating these priors to deep features, AdaptSplat injects them into the Query-Key similarity space of attention: $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$6 Only $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$7 and $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$8 are modified; $$\Sigma = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top},$$9 is left unchanged (Xing et al., 11 May 2026). The paper describes this as a high-frequency positional encoding. It is not positional encoding in the standard sinusoidal sense; rather, it is a spatially aligned, image-conditioned structural bias injected into similarity computation.

On the decoder side, FPA generates multi-scale adaptive residual modulation masks. The modulation rule is

$$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$0

where $$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$1 is generated by a Sigmoid and $$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$2 is learnable (Xing et al., 11 May 2026). The effect is spatially selective amplification of shallow residual detail during decoding. In flat regions the residual path remains weakly modulated; in high-frequency regions the shallow feature contribution is amplified.

Training is end-to-end with

$$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$3

and

$$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$4

The paper explicitly states that Focal Frequency Loss, $$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$5, creates synergy with FPA by emphasizing high-frequency components in the frequency domain (Xing et al., 11 May 2026).

4. Relationship to adjacent frequency-aware adapter families

Several recent methods occupy nearby design space but should not be conflated with FPA. They differ chiefly in what is conditioned, what is decomposed, and whether the goal is preservation, alignment, or selection.

Method	Domain	Closest characterization
FeRA (Yin et al., 22 Nov 2025)	Diffusion fine-tuning	Frequency-aware, frequency-aligned LoRA expert routing
FAA (Bae et al., 26 Dec 2025)	LLM PEFT	Fourier-Activated Adapter with RFF-based decomposition and adaptive gating
FSAM frequency adapter (Bui et al., 11 May 2026)	Medical segmentation	Frequency-aware feature injection using Fourier amplitude
Earth-Adapter (Hu et al., 8 Apr 2025)	Remote-sensing segmentation	Mixture of frequency adaptation with spatial/LF/HF experts
VMDA visual adapter (Xu et al., 30 Jun 2025)	Multi-modal tracking	Frequency-guided fusion via learned high/low reweighting

FeRA is built around a coarse-to-fine frequency-energy progression in diffusion denoising. Its Frequency-Energy Indicator, soft routing, and Frequency-Energy Consistency Loss are designed to align adapter corrections with the current residual spectral demand; the paper explicitly states that calling it frequency-preserving would be more interpretive than literal (Yin et al., 22 Nov 2025). FAA injects random Fourier features into adapter paths and learns channel-wise and hierarchical gating; it is a frequency-aware activation mechanism rather than an identity-preserving spectral branch (Bae et al., 26 Dec 2025). FSAM passes Fourier-amplitude representations through patch embedding and lightweight adapters attached to SAM’s ViT encoder, but it uses only amplitude and introduces no explicit frequency-preservation regularizer (Bui et al., 11 May 2026). Earth-Adapter uses DFT-based low-/high-frequency decomposition with spatial, LF, and HF experts plus a router, yet its own formulation is best described as frequency-aware selective adaptation rather than strict preservation (Hu et al., 8 Apr 2025). VMDA uses a “frequency selector,” but its decomposition is implemented with pooling, convolution, normalization, and gating rather than FFT, DCT, or wavelets, and the mechanism is aimed at selective enhancement and cross-modal transfer (Xu et al., 30 Jun 2025).

The distinctive feature of FPA, by contrast, is that high-frequency priors are extracted from shallow features and then injected specifically into attention similarity and multi-scale residual reconstruction so as to counteract high-frequency attenuation in a 3DGS pipeline (Xing et al., 11 May 2026).

5. Empirical evidence

The most direct evidence for FPA comes from AdaptSplat’s ablations on a DL3DV subset with $$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$6k scenes and $$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$7k iterations (Xing et al., 11 May 2026). Starting from a generic baseline, the reported progression is:

• Baseline: PSNR $$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$8, LPIPS $$\Sigma' = \mathbf{J}\mathbf{W}\Sigma\mathbf{W}^{\top}\mathbf{J}^{\top}.$$9, SSIM $s_x \approx s_y \approx s_z,$0
• + VFM: PSNR $s_x \approx s_y \approx s_z,$1, LPIPS $s_x \approx s_y \approx s_z,$2, SSIM $s_x \approx s_y \approx s_z,$3
• + FPA: PSNR $s_x \approx s_y \approx s_z,$4, LPIPS $s_x \approx s_y \approx s_z,$5, SSIM $s_x \approx s_y \approx s_z,$6
• + FFL: PSNR $s_x \approx s_y \approx s_z,$7, LPIPS $s_x \approx s_y \approx s_z,$8, SSIM $s_x \approx s_y \approx s_z,$9
• + M-FPA: PSNR $1.5$0, LPIPS $1.5$1, SSIM $1.5$2

These numbers support two claims made by the paper: first, attention-side FPA improves over the VFM-equipped baseline; second, extending FPA to multi-scale decoder modulation yields an additional gain (Xing et al., 11 May 2026).

The prior-extraction ablation is especially informative. Using the same ablation subset, the paper reports:

• Base: $1.5$3
• Fourier: $1.5$4
• Conv: $1.5$5
• Sobel: $1.5$6
• Ours: $1.5$7

Here the wavelet-based extractor substantially outperforms the Fourier and learnable high-pass convolution alternatives and also exceeds Sobel in SSIM and LPIPS (Xing et al., 11 May 2026). This supports the paper’s claim that orthogonal scale-frequency decomposition provides a more useful prior than generic spectral filtering or simple edge detection.

The paper also analyzes Gaussian anisotropy on RE10K using Fractional Anisotropy. The reported values are:

• w/o FPA: FA $1.5$8
• w/ FPA: FA $1.5$9

Higher FA is presented as evidence that FPA reduces isotropic degeneration and encourages more anisotropic Gaussians aligned with structural boundaries (Xing et al., 11 May 2026). Qualitatively, the paper reports sharper geometric boundaries, better handling of overlapping glassware, clearer local tabletop textures, and reduced blurring and structural degradation.

6. Interpretation, limitations, and historical placement

FPA is best understood as a preservation-oriented adapter for high-frequency structural priors, not as a guarantee of exact spectral invariance (Xing et al., 11 May 2026). The method preserves geometry-relevant detail by recovering direction-aware high-frequency evidence from shallow features and reintroducing it where deep architectures are most prone to smoothing: attention affinity computation and decoder upsampling. A plausible implication is that “frequency-preserving” here should be read as preserving access to high-frequency structure during adaptation and reconstruction, rather than preserving the entire input spectrum unchanged.

The formulation also has practical limits. The paper does not provide an exact layer-by-layer breakdown of FPA, does not specify the precise DWT wavelet family, does not give exact channel dimensions for $\mathbf{F}_{hf}$0, and does not fully detail any projection layers used after DWT (Xing et al., 11 May 2026). Some quantities, such as the Fractional Anisotropy expression, are reported with formatting issues. The design is demonstrated with a hierarchical DINOv3-ConvNeXt backbone and is therefore most concretely specified for that setting.

Historically, FPA should be situated within a broader movement toward frequency-aware PEFT and adapter design, but it remains more specific than those neighboring approaches. FeRA centers frequency-energy-conditioned routing in diffusion, FAA centers random Fourier features and adaptive gating in LLMs, FSAM injects Fourier amplitude into SAM for medical segmentation, Earth-Adapter decomposes remote-sensing features into spatial/LF/HF branches, and VMDA uses shallow high/low reweighting for multi-modal tracking (Yin et al., 22 Nov 2025, Bae et al., 26 Dec 2025, Bui et al., 11 May 2026, Hu et al., 8 Apr 2025, Xu et al., 30 Jun 2025). By contrast, FPA is a 3DGS-specific design whose stated purpose is to preserve high-frequency geometric fidelity against over-smoothing.

Finally, the acronym itself requires caution. In CenterFPANet, FPA denotes an FPN-plus-attention module for optical remote sensing object detection and is unrelated to frequency-domain preservation or adapters in the modern PEFT sense (Gu et al., 2020). Within contemporary adapter terminology, the technically precise meaning of Frequency-Preserving Adapter therefore depends on context, and in the most explicit recent usage it refers to the AdaptSplat module built around shallow wavelet priors, Query-Key structural biasing, and adaptive residual modulation (Xing et al., 11 May 2026).