1 Introduction↩︎

Feed-forward models for 3D reconstructions [1]–[11] generate 3D Gaussian representations [12] directly from input views. These approaches predict pixel-aligned Gaussians, where each input pixel produces the parameters of a corresponding 3D Gaussian in the scene. While this design enables efficient feed-forward inference, it allocates Gaussians uniformly across the image without considering the underlying scene complexity. A straightforward way to reduce this redundancy is to prune the Gaussians. However, naive pruning strategies introduce severe artifacts in rendered novel views, often producing empty regions that disrupt scene continuity and degrade rendering quality.

Several works have explored compression techniques for 3D Gaussian representations [13]–[18]. However, these approaches primarily focus on post-hoc quantization or encoding and typically require dense multi-view inputs along with scene-specific optimization. Such reliance on iterative optimization limits their applicability to the challenging feed-forward setting. Other works attempt to prune pixel-aligned Gaussians directly within feed-forward pipelines using geometry-driven heuristics such as overlap removal [7], opacity thresholding [9], or enforcing a single Gaussian per voxel [3]. While computationally simple, these strategies are largely agnostic to scene complexity and offer little control over the pruned Gaussians. Consequently, they may remove primitives that are necessary to represent high-frequency structures [10], [19]. These limitations raise an important question: Can we enable controllable 3D Gaussian density in feed-forward models without sacrificing their fidelity?

We draw inspiration from well-established statistical properties of natural images, whose power spectral density (PSD) follows an approximate power-law decay (\(PSD \propto \frac{1}{f^\alpha}\), where \(f\) denotes spatial frequency and \(\alpha \approx 2\)) [20]. This observation implies that most visual energy is concentrated in low-frequency components, while high-frequency structures are sparse and localized. Consequently, different regions of a scene require different levels of representation density: smooth areas with low-frequency content can be reconstructed with relatively few Gaussians, whereas highly textured regions containing fine structural details require denser Gaussian allocation. Prior work [10], [19] empirically supports this hypothesis, demonstrating that high-frequency regions benefit from increased Gaussian density for accurate reconstruction. However, existing feed-forward architectures largely ignore this requirement due to their fixed pixel-aligned prediction strategy. To further illustrate this observation, we conduct a simple experiment shown in Fig. 2.

Motivated by this key insight, we propose AdaptiveSplat, a feed-forward framework that enables controllable 3D Gaussian allocation while preserving reconstruction fidelity. Our method focuses on capturing the variance in visual complexity. This is achieved by estimating the texture of the scene. Based on this signal, we perform texture-aware pruning that removes redundant Gaussians more aggressively in smooth regions while preserving those necessary for detailed structures. To compensate for the removed primitives, we introduce a novel adaptive Gaussian head that re-predicts the attributes of the remaining Gaussians within the same feed-forward pipeline. This design enables flexible control over the Gaussian budget without requiring test-time optimization or post-pruning fine-tuning. Importantly, our approach can be seamlessly integrated into existing feed-forward 3D Gaussian pipelines without requiring architectural modifications. In summary, our contributions are as follows:

• Texture-Based Scene Energy Estimation: We introduce a texture-driven scene energy estimation mechanism that quantifies the amount of high-frequency content present in different image regions, enabling identification of areas that require higher representational capacity.

• Texture-Aware Selective Pruning: Building on the estimated scene energy, we propose a principled pruning strategy that selectively removes Gaussians from low-texture regions while preserving those in high-detail areas.

• Adaptive Gaussian Head: To mitigate artifacts introduced by pruning, we design an adaptive Gaussian head that re-predicts the attributes of the retained Gaussians, allowing the representation to adjust to the modified Gaussians.

• We perform extensive experiments for various feed-forward models on datasets like RE10K, ACID, DL3DV, Tanks and Temples, and DTU.

2 Related Work↩︎

2.1 Radiance Fields and Multi View Reconstruction↩︎

Reconstructing 3D scenes from multi-view images is a core computer vision problem. Modern radiance field methods like NeRF [21], [22] and 3D Gaussian Splatting (3DGS) [12], [23] achieve high-quality novel view synthesis, with 3DGS offering notable advantages in rendering speed through rasterization.[24] and [25] use structural anchors and contextual compression to reduce redundancy. These approaches have found success in applications such as 3D asset generation [26]–[33], scene editing [34]–[42], and dynamic scene reconstruction [12], [43]–[46]. However, they often require dense multi-view images which limit adaptability and efficiency in dynamic, real-time environments. Sparse-view reconstruction methods [4], [6], [26], [47]–[50] leverage learned priors from datasets or foundational models [51]. Recently, Feed-forward methods [52]–[56] offer a compelling alternative, directly inferring 3D representations with superior generalization across diverse scenes, bypassing the computationally expensive steps of Structure-from-Motion (SfM) [57] and radiance field optimization by directly predicting dense 3D scene representations from limited views [19], [58]–[60] but are rigid in number of 3D Gaussian predictions. These methods have been extended to dynamic scenes [61], multi-view consistency [62], and estimating 3DGS representations [1], [2], [10], [63]–[65], even from just a few views [3], [9].

2.2 Compression in 3D Gaussian Splatting↩︎

Due to its explicit nature, 3DGS often requires a large number of Gaussians to represent a scene, leading to higher storage, processing time, and slow rendering. To address these challenges, several works focus on compressing 3DGS representations. EAGLES [13] prunes based on scene-wide influence, targeting low-contribution Gaussians (e.g., small scale, low opacity, occlusion) and LightGaussian [66] prunes based on estimated importance (ray hits, volume), PUP-3DGS [67] uses uncertainty estimation to estimate Gaussians’ importance by approximating second order gradient of loss. EfficientGS[68] estimates importance of Gaussians per pixel by selecting Top-K Gaussians to be retained, and SOG-3DGS [14] applies image compression techniques. Other methods like [16] remove Gaussians with low gradients followed by fine-tuning, while [15] proposes pruning and densification based on alpha blending importance. However, a common limitation across these compression techniques is their reliance on quantization or their inability to dynamically modify the attributes of remaining Gaussians post-pruning, which can lead to visible artifacts. Moreover, all existing methods require dense multi-view images of the scene to compress which fail in the sparse-view feed-forward setting. Interestingly, previous work like [69] has explored using wavelet based frequency decomposition to achieve high quality synthesis for multiview images but no such method exists for 3D Gaussian Splatting representation.

3 Method↩︎

We introduce a controllable feed forward 3D Gaussian generation pipeline. 3 shows the overall pipeline of our method.

3.0.0.1 Problem Formulation

Existing feed-forward 3D Gaussian prediction models generate a pixel-aligned set of 3D Gaussians \(\mathcal{G}_k\) for each input view. Given \(m\) views \(\{(I_k, v_k)\}_{k=1}^{m}\), where \(I_k \in \mathbb{R}^{H \times W \times 3}\) denotes the RGB image and \(v_k\) its associated camera pose, such models produce a set of \(N = mHW\) primitives without explicit control over representation size. In contrast, we introduce explicit control over the number of retained Gaussians. Under a user-defined pruning budget \(\beta \in [0,1)\), our objective is to predict a compact subset \(\mathcal{G} = \{g_i\}_{i=1}^{B}\) where \(B=\lfloor(1-\beta)N\rfloor\). Each Gaussian is parameterized as \(g_i = \{\boldsymbol{\mu}_i, \boldsymbol{s}_i, \boldsymbol{q}_i, \boldsymbol{\alpha_i}, \boldsymbol{\rho}_i\}\), where \(\boldsymbol{\mu}_i \in \mathbb{R}^3\) denotes the 3D mean, \(\boldsymbol{s}_i \in \mathbb{R}^3\) anisotropic scale, \(\boldsymbol{q}_i \in \mathbb{R}^4\) a unit quaternion encoding rotation, \(\alpha_i \in [0,1]\) opacity, and \(\boldsymbol{\rho}_i \in \mathbb{R}^{h}\) spherical harmonics coefficients modeling view-dependent appearance.

3.1 Texture-Based Scene Energy Estimation↩︎

We estimate spatially localized scene complexity using the Discrete Wavelet Transform (DWT) applied to the input images. Given images \(\{I_k\}_{k=1}^m\), its single-level 2D DWT is computed as \(W_{ij}^{d} = \sum_{u}\sum_{v} I_k(u,v) \cdot \psi^{d}(i-u, j-v)\), where \(\psi^{d}\) denotes the directional wavelet basis. This decomposes the image into three high-frequency detail components corresponding to horizontal, vertical, and diagonal orientations. We define the local texture energy at pixel \((i,j)\) in view \(k\) as \(E_{i,j}^{(k)} = \sum_{d \in {h,v,diag}} |W_{ij}^{d}|\), which aggregates directional high-frequency responses.

3.2 Texture-Aware Selective Pruning↩︎

Using a pretrained multi-view backbone, we extract a dense pointmap \(\mathcal{P}_k = \{ p_i\}_{i=1}^{HW}\) from each input view \(I_k\) where each point \(p \in \mathbb{R}^6\) encodes its 3D location and RGB color. We aggregate all views to form a global point cloud \(\mathcal{P} = \bigcup_{k=1}^{m} \mathcal{P}_k\), where \(\mathcal{P} = \{ p_i\}_{i=1}^{N}\). To obtain homogeneous regions, we partition \(\mathcal{P}\) into \(K\) SuperClusters using \(K\)-means clustering. Clustering in \((x,y,z,r,g,b)\) space encourages points that are spatially proximate and photometrically consistent to be grouped together. Such regions correspond to locally smooth surfaces that can be represented with fewer but spatially broader Gaussians. Formally, each SuperCluster \(SC_i\) is defined as \(SC_i = \{ p \in \mathcal{P} \mid |p - c_i|^2 \leq |p - c_j|^2, \forall j \neq i \}\) where the centroid is \(c_i = \frac{1}{|SC_i|} \sum_{p \in SC_i} p\). The resulting clusters \(\{SC_i\}_{i=1}^{K}\) form a disjoint partition of the point cloud such that \(\bigcup_{i=1}^{K} SC_i = \mathcal{P}\) and \(SC_i \cap SC_j = \emptyset\) for \(i \neq j\).

We assign texture energy to each SuperCluster by averaging the wavelet coefficients associated with its constituent points \(p\). Let \(\pi_k(p)\) denote the projection of point \(p\) onto image \(I_k\) via pointmap \(\mathcal{P}_k\). The SuperCluster energy is computed as \(E_{SC_i} = \frac{1}{|SC_i|} \sum_{p \in SC_i} \bar{E}(\pi(p))\) where \(\bar{E}(p)\) denotes the wavelet coefficient of the pixel corresponding to the point \(p\). These SuperCluster-level energies provide a structured measure of regional complexity and guide subsequent Gaussian allocation under the pruning budget.

3.2.0.1 Binary Mask Construction:

\(\{SC_1,\dots,SC_K\}\) denote the regions ordered by increasing energy \(E_{SC_j}\). Each region \(SC_j\) has a nominal retention of \(\gamma |SC_j|\) Gaussians, where \(\gamma=0.1\) (empirically validated, ablation provided). We choose the largest index \(\ell\) such that \(\sum_{j=1}^{\ell} \gamma |SC_j| \;\le\; B\). This ensures that we can effectively cover the entire low-texture region with 3D Gaussians with fewer retained Gaussians while pruning the rest.

For all regions \(j<\ell\), we retain the \(\gamma |SC_j|\) Gaussians. For the final region \(SC_\ell\), we retain only \(|R_\ell| = B - \sum_{j=1}^{\ell-1} \gamma |SC_j|.\) Let \(\mathcal{G}_{\mathrm{ret}}\) denote the retained set of Gaussians. The corresponding binary mask entries for each context view \(i \in \{1,\dots, m\}\) are given by \(M_{i,j} = \mathbb{1}\!\left[\, G_j \in \mathcal{G}_{\mathrm{ret}} \cap \mathcal{G}_i \,\right], \quad j \in \{1,\dots,HW\}\) (see Fig. 4).

3.3 Adaptive Gaussian Head↩︎

Given the features \(F\) and binary masks \(M_1, \dots M_m\), the Adaptive Gaussian Head \(f_\theta\) predicts refined Gaussian attributes for the retained indices:

\(f_\theta(F,M) \rightarrow \{\boldsymbol{s}_i, \boldsymbol{q}_i, \boldsymbol{\alpha_i}, \boldsymbol{\rho}_i\}_{i=1}^{B}\). The Adaptive Gaussian Head is built on the DPT architecture [70] with additional convolutional layers that fuse ViT [71] features extracted from the multiview feature map \(F\) with binary masks \(M_1, \dots, M_m\).

3.4 Training and Inference↩︎

3.4.0.1 Training:

To ensure that the model’s performance is stable across diverse budget rates \(\beta\) at test time, it is important that it generalizes well to a broad spectrum of pruning percentages during training. To achieve this, we train the Gaussian Head with \(\beta \sim \mathcal{U}[0, 1)\). This improves robustness of the method to different budgets \(\beta\), enabling adaptation of the retained 3D Gaussians.

The network is trained using photometric loss on the rendered image from the retained 3D Gaussians at the target camera pose and the corresponding ground truth image using the following loss formulation \(\mathcal{L}\) as follows: \[\mathcal{L} = \frac{1}{|v_t|} \sum_{v \in {v_t}} \left( \|I_{v} - \hat{I}_{v}\|_2^2 + \lambda \cdot \mathrm{LPIPS} \left(I_{v}, \hat{I}_{v}\right) \right),\] Here \(v_t\) represents the set of target camera views, \(I_v\) denotes the target ground truth image and \(\hat{I}_v\) denotes the rendered image from the target view \(v\). 3D Gaussians are obtained from Gaussian Head \(f_\theta\) using context view decoder features along with its corresponding masks \(M_1, M_2, \dots M_m\). The weighting coefficient \(\lambda\) controls the perceptual regularization strength.

3.4.0.2 Inference:

At inference, the user selects a desired sparsity level \(\beta \in [0 ,0.8]\). The system computes DWT from images, forms masks, and directly predicts adapted Gaussian attributes via the Gaussian Head.

4 Experiments↩︎

4.1 Experimental Setup↩︎

4.1.0.1 Implementation Details

Our approach builds upon the architecture of MASt3R [53] for sparse-view inputs and VGGT [54] for dense-views with a DPT [70] Gaussian Head to predict 3D Gaussian primitives similar to [2], [3], which integrates a Gaussian prediction head into the MASt3R/VGGT pipeline. This allows for extraction of 3D Gaussian attributes for each point in the point cloud, providing an initial 3D representation. These methods don’t require camera poses as input; thus, we also implement our approach on MVSplat[5] backbone which also takes camera pose as input. We choose \(K=300\) and \(\gamma=10\%\) for our pipeline (ablations provided on these hyperparameters) and \(\lambda\) is set to \(0.001\). We trained our model on 1 NVIDIA A100 40GB GPU for 48 GPU hours on each dataset. We use available pretrained checkpoints of the baseline feed-forward models for the evaluation.

4.1.0.2 Datasets

We train and evaluate our method on diverse real-world datasets following the experimental setup of pixelSplat [4] and AnySplat [3]. Our primary training and evaluation data is derived from RealEstate10K [72] (indoor, multi-view videos), ACID [73] (large-scale, dynamic outdoor drone scenes) and DL3DV [74] (large scale scene) datasets. These datasets provide multi-view images with COLMAP-computed [57] camera poses, and we adhere to their official train-test splits. To assess zero-shot generalization, we also evaluate on the DTU dataset [75], [76], comprising high-quality, object-centric scans. We further report results on the Tanks and Temples dataset [77] in the supplementary material.

4.1.0.3 Evaluation Criteria

We benchmark our method on novel view synthesis using widely adopted image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [78], and Learned Perceptual Image Patch Similarity (LPIPS) [79]. Beyond these measures, we explicitly compare against \(\beta\) used for 3D scene representation. This metric directly demonstrates our approach’s efficiency in reducing redundancy while maintaining superior visual fidelity.

4.2 Comparisons with Baselines↩︎

Comparisons with existing feed-forward methods are challenging because they lack controllable Gaussian allocation and are designed for dense, per-pixel prediction, unlike our approach. To enable fair comparison, we construct baselines by applying established pruning techniques on top of state-of-the-art feed-forward methods. We consider four representative sparse-view feed-forward methods: pixelSplat [4], MVSplat [5], NoPoSplat [2], and HiSplat [10], as well as Gaussian Graph Network (GGN) [7], AnySplat [3], a recent method designed for dense-view settings. On top of the Gaussian primitives predicted by these models, we apply four representative pruning strategies: EAGLES [13], LightGaussian [66], EfficientGS [68], and PUP3DGS [67].

To compare with existing methods fairly, baselines must also operate under the feed-forward setting. However, directly applying pruning to the predicted Gaussians is suboptimal, as most pruning strategies assume per-scene optimization. While this optimization can partially recover reconstruction quality, it breaks the feed-forward assumption. To account for this trade-off, we evaluate baselines under two settings: (1) pruning followed by per-scene fine-tuning, which we refer to as with fine-tuning, and (2) direct pruning, which preserves the feed-forward property and is referred to as without fine-tuning. We evaluate our approach in both sparse-view and dense-view feed-forward settings. For sparse-view evaluation, we use two input views, following the protocol of pixelSplat [4]. For dense-view settings we evaluate using 6, 9, 16, and 32 view inputs following the protocol of AnySplat [3]. We report results for the best-performing backbone, while detailed comparisons across all backbones pixelSplat, MVSplat, HiSplat, NoPoSplat, GGN, InstantSplat [80] and AnySplat are provided in the supplementary.

4.3 Inference Time Analysis↩︎

To highlight the practical benefits of controllable pruning, we report rendering throughput under different pruning strengths. We measure rendering speed by synthesizing \(1000\) novel views for each pruning strength and reporting the Frames Per Second (FPS).

4.4 Ablation Study↩︎

We analyze the impact of the key design choices in our method. Each ablation isolates a specific component to evaluate its contribution to the overall performance.

\(\bullet\) Texture Ablation: We analyse the importance of texture for ranking SuperClusters. In our method, texture energy determines the order in which SuperClusters are selected for pruning. In this ablation, we remove the texture-based ranking and train the model end-to-end while selecting SuperClusters randomly for pruning to evaluate the importance of texture guidance.

\(\bullet\) Importance of Hyperparameter \(K\) and \(\gamma\): We provide ablations on various choices of \(K\) and \(\gamma\), which control the number of SuperClusters and the retention ratio within each SuperCluster.

\(\bullet\) Adaptive vs Fixed Gaussian Head: Here, we replace the Adaptive Gaussian Head with a fixed, pretrained Gaussian head and directly select the retained Gaussians from its predictions without further adaptation. This allows us to evaluate the importance of learning adaptive Gaussian attributes after pruning.

\(\bullet\) Binary Mask: To validate the importance of the binary mask, we train the Adaptive Gaussian Head with the mask as input and compare it with another Adaptive Gaussian head trained end-to-end without the mask.

Collectively, these ablations validate the importance of the key components of our method in achieving superior performance.

4.5 Qualitative Texture Analysis↩︎

The experiments so far focus on novel view synthesis to evaluate reconstruction quality. However, our method explicitly leverages texture signals to guide Gaussian allocation. To analyze this component more directly, we perform qualitative analysis with baselines focusing on the trends in the texture structure of the reconstructed scenes.

5 Results↩︎

3pt

3pt

5.1 Comparison with Baselines↩︎

5.1.0.1 Quantitative Analysis

Tab. 1 presents a detailed comparison of AdaptiveSplat against established baselines on the ACID and RE10K datasets with finetuning for sparse view input (additional evaluations without finetuning included in supplementary) and Tab. 2 presents results on the DL3DV dataset with finetuning on dense view (additional evaluations without finetuning included in supplementary). Performance is reported across \(\beta=\{ 0.4, 0.6, 0.8\}\). Our method consistently outperforms existing pruning strategies across all datasets and \(\beta\) values. On ACID, AdaptiveSplat improves PSNR from \(19.43dB\) (EAGLES) to \(22.55dB\) at \(\beta=0.4\), while reducing LPIPS from \(0.620\) to \(0.299\). On RE10K, it achieves \(22.29\) PSNR compared to \(16.60\) from LightGaussian and \(16.53\) from EAGLES. Improvements are even larger on DL3DV, for example with \(6\) views and \(\beta=0.4\), AdaptiveSplat reaches \(20.45dB\) PSNR versus \(11.36dB\) from LightGaussian. Notably, while baseline methods degrade significantly at high pruning (\(\beta=0.8\)), our method maintains strong performance across all view counts.

Overall, the trend remains the same, highlighting the superiority of the proposed approach. We report only the best-performing baselines here, while the remaining baseline comparisons, wider pruning ranges, ablation studies (with and without fine-tuning across various backbones), and cross-domain evaluation on the DTU dataset are deferred to the supplementary material.

5.1.0.2 Qualitative Analysis

We present a qualitative comparison of our method in Fig. 5 and Fig. 6. In all these results, we highlight two important issues with the baseline methods: (a) For non-finetuned cases, “black patch" artifacts can be consistently seen, which is indicative of empty spaces within the 3D scene where Gaussians were removed without any readjustment. Because of”black patches" artifacts, we see lower metrics across all baselines compared, whereas our results show consistent results without any such artifacts and simultaneously provide high fidelity. (b) For finetuned cases, results struggle with “blobby Gaussian” artifact. This happens as existing pruning methods completely remove all Gaussians from a local region without replacing them. To compensate for the empty space, Gaussians from neighboring areas must expand, which distorts their shape and negatively impacts the quality of the overall representation. This effect persists even after further finetuning, as the fundamental problem of sub-optimal Gaussian allocation remains.

We also show effects of increasing \(\beta\) in Fig. 6, where we pick three broad pruning rates (\(\beta= \{0.4, 0.6, 0.8 \}\)) and compare AdaptiveSplat with the baselines. As can be seen, baselines degrade the scene quality and completely fail to adapt existing Gaussian attributes at higher \(\beta\) value. In contrast, AdaptiveSplat optimally re-adjusts 3D Gaussians to adapt to higher pruning rates, consistently maintaining the high scene quality. We show more qualitative results in the supplementary.

5.2 Ablation Study↩︎

3 evaluates the contribution of key components of our method at a fixed pruning budget \(\beta=0.4\). Our analysis reveals that removing the Adaptive GS Head training significantly degrades performance, reducing PSNR from \(22.55\) dB to \(17.44\) dB on ACID and from \(22.29\) dB to \(19.81\) dB on RE10K, which indicates that learning to adapt Gaussian attributes post-pruning is critical for recovering reconstruction quality. Similarly, excluding the DWT-based texture signal leads to noticeable performance decreases specifically from \(22.55\) dB to \(21.31\) dB on ACID confirming the importance of texture guidance in ranking SuperClusters and allocating Gaussians effectively. Finally, removing the contextual mask further degrades RE10K performance to \(20.06\) dB, highlighting its vital role in providing spatial context for the Adaptive Gaussian Head. The impact of hyperparameters \(K\) and \(\gamma\) is evaluated in 4, where increasing \(K\) generally improves reconstruction quality by producing finer SuperClusters that better identify low-texture regions and enable more precise allocation of larger Gaussians. However, these quality gains become marginal beyond \(K=300\) while training costs continue to rise. Regarding the retention ratio, we observe that retaining \(\gamma=0.10\) of Gaussians achieves the best performance, particularly under high pruning scenarios such as \(\beta=0.9\). Larger \(\gamma\) values tend to introduce redundant Gaussians in the same regions, which ultimately degrades reconstruction quality, consistent with observations in prior works [13], [15], [66]. Overall, these results demonstrate that each individual component contributes significantly to the final performance, with their combination yielding the highest reconstruction quality across both the ACID and RE10K datasets.

3pt

4pt

5.3 Qualitative Texture Analysis↩︎

We provide additional analysis focusing on scene texture preservation. Our method explicitly targets low-texture regions for pruning, allowing the model to allocate fewer but larger Gaussians in such areas while preserving fine details in highly textured regions.  ¿tbl:fig:anysplat959view95results? presents a qualitative comparison of texture preservation on a representative scene across different pruning ratios. We observe that baseline methods tend to oversmooth textured regions under pruning, leading to noticeable loss of high-frequency details. In comparison, our approach better maintains scene fidelity by leveraging texture-aware clustering and natural scene statistics. Furthermore, Fig. [tbl:fig:anysplat95texture95results] provides additional qualitative comparisons across multiple scenes at \(\beta = 0.4\), demonstrating that while baselines suffer from blurred edges, over-smoothing, and structural distortion, our method maintains sharp high-frequency details and structural integrity.

These results consistently demonstrate that competing methods introduce excessive smoothing in high-texture areas, whereas our method preserves structural details more effectively under increasing pruning strengths. ¿tbl:fig:scale95statistics? shows the histogram of mean scale values demonstrating our insight empirically that mean scale values of 3D Gaussians corresponding to low texture regions are predicted larger in our method to retain the quality of scene; thus, redundant 3D Gaussians can be pruned.

5.4 Inference Time Analysis↩︎

6 Conclusions and Future Work↩︎

We introduce AdaptiveSplat, a novel, feed-forward framework for controllable 3D Gaussian pruning. Our method’s core novelty is texture-aware Gaussian control that significantly reduces primitives by targeting low-texture regions. By eliminating the need for expensive per-scene optimization, AdaptiveSplat provides a practical solution for controllable 3D reconstruction, achieving a superior quality-efficiency trade-off even at extreme pruning rates. On smooth surfaces with high-frequency texture (such as intricate wall patterns) or specular highlights, our method suffers from the same limitation as any pixel-aligned feed-forward model. We briefly discuss these limitations in the supplementary. Future work will focus on integrating a more nuanced understanding of both texture and geometry to enable more robust pruning decisions, paving the way for truly efficient 3D scene representation.

Acknowledgement↩︎

We sincerely thank Anudeep for careful review of our manuscript and for providing valuable feedback and suggestions that significantly improved the quality of the article. This work is supported by Google and Kotak IISc AI-ML Centre (KIAC). Badrinath Singhal is supported by TCS Research Scholarship and Qualcomm Innovation Fellowship (QIF).

Supplementary Material: AdaptiveSplat: Texture Aware Controllable 3D Gaussian Allocation for Feed-Forward Reconstruction↩︎

Table of Contents↩︎

7 Additional Baselines↩︎

We implemented additional baselines where pruning masks are first generated using existing pruning pipelines and then concatenated with the Gaussian Head input to predict the attributes of the retained Gaussians. The Gaussian Head for these baselines is trained following the same procedure and loss functions as in our method to ensure fairness.

We evaluate these baselines on both the ACID and RE10K datasets, and the results are provided in Tables 5 and 6. These enhanced baselines achieve stronger results than the post-hoc pruning baselines reported in the main paper, particularly at higher pruning ratios. However, our method still consistently outperforms them across all pruning levels and datasets, primarily due to our texture-aware region selection strategy, which better preserves scene structure under aggressive pruning.

It is also important to note that these additional baselines are not feed-forward: they require (1) predicting the full set of 3D Gaussians, (2) generating a pruning mask, and only then (3) predicting the attributes of the retained Gaussians in a second pass. This makes them inherently two-stage. In contrast, our approach performs pruning within the feed-forward generation pipeline, offering controllable pruning during inference.

2pt

MASt3R and VGGT are foundation models that generalize well, so we built on AnySplat and NoPoSplat, which use them as backbones. Our method is pose-agnostic; to demonstrate this, we further extend it to MVSplat, a posed feed-forward method, on ACID dataset. Vanilla MVSplat returns a fixed Gaussian count (\(\beta=0\)). We evaluate our approach over \(\beta\in\{0,0.4,0.6,0.8\}\) (7), confirming that it works across both posed and unposed feed-forward methods.

8 Extension to Dense View input↩︎

We additionally evaluate our texture-aware pruning strategy on the InstantSplat framework. We used Tanks and Temples dataset for our experiment, specifically we used the ‘Horse’ scene from Tanks and Temples. We first use MASt3R to obtain the pixel-aligned point cloud from sparse-view images, construct texture-based SuperClusters, and rank them using our wavelet textureness metric. We prune points according to the desired pruning strength and then train the 3DGS stage for 1000 iterations. Importantly, we disable densification and pruning within InstantSplat to ensure that the Gaussian count remains fixed throughout the optimization pipeline.

We report results for 3, 4, and 10-view input settings and evaluate on novel views (Tab.  8). Our method consistently achieves higher PSNR, SSIM, and lower LPIPS across all pruning strengths and view configurations. This demonstrates that the proposed texture-aware pruning not only preserves reconstruction quality but also enhances efficiency in multiview sparse-input pipelines without requiring any architectural changes to InstantSplat.

Additionally we provide results for dense view feed forward in 9 on DL3DV where we do not finetune the remaining Gaussians. The results are provided for 6, 9, 16 and 32 input views using VGGT as multiview feature extractor. The corresponding results for fine tune are included in main paper. 10 reports the pretrained AnySplat baseline against Ours at \(\beta=0\) on DL3DV across 9, 16, and 32 input views.

10pt

9 Inference Latency↩︎

To highlight the practical benefits of controllable pruning, we report rendering throughput under different pruning strengths. We adopt the InstantSplat setting for sparse-view reconstruction, where MASt3R provides the initial pixel-aligned points and the 3D Gaussian attributes are subsequently optimized. To ensure a fair comparison, we disable densification so that the number of Gaussians remains fixed to the MASt3R initialization. We used the Tanks and Temples dataset specifically the Horse scene for this analysis.

We measure rendering speed by synthesizing 1000 novel views for each pruning stage and reporting the Frames Per Second (FPS). We utilize our pruning technique to prune the Gaussians. Tab. 11 summarizes the FPS achieved across pruning strengths, and Fig. ¿tbl:fig:fps-prune-plot? visualizes this trend. As pruning strength increases, the reduced number of active Gaussians consistently improves rendering speed, demonstrating the effectiveness of pruning as a controllable knob for balancing reconstruction quality and runtime efficiency. [tbl:fig:pareto] presents pareto plot to compare inference time and quality analysis of ours with other methods.

We report the end-to-end latency of predicting 3D Gaussians from sparse input views. Tab.  12 summarizes the runtime across different baselines. Our method exhibits latency comparable to the non-finetuned baseline and is substantially faster than baselines that require per-scene finetuning of the remaining Gaussians. This highlights that AdaptiveSplat retains the feed-forward efficiency of MASt3R while enabling controllable pruning without the overhead of optimization. Since rendering is efficient, changing \(\beta\) does not significantly affect the result in our method.

10 Results with naive pruning baselines↩︎

We include results with naive pruning baselines for completeness. We include results with pruning based on opacity as well as low resolution input to decrease number of 3D Gaussian primitives. Specifically we remove lowest opacity Gaussians till pruning budget \(\beta\) is met; similarly, we downscale the image till the total number of pixels in multi view input images meets the \(\beta\) budget. Tab. 13 presents these results, we observe that this naive pruning degrades the scene as the remaining Gaussians are not refined to adjust for the removed primitives.

4pt

11 Results on different backbones↩︎

Here we present detailed results of our experiments on RE10K and ACID dataset for different pruning stages on NoPoSplat and HiSplat backbone. 14, 15 presents results without finetuning on both ACID and RE10K, respectively, and 8 demonstrates results for ACID and RE10K for both with and without finetuning. We also include results with different feed-forward backbones with various pruning strategies. Specifically we use pixelSplat and MVSplat backbones on RE10K and ACID dataset with all pruning strategies discussed before (we’ve also added random 3D Gaussian pruning) along with before and after finetuning results. 16 on RE10K dataset and 17 on ACID dataset presents results of all pruning stages for 2 different feed-forward backbone on 4 different pruning strategies where we don’t fine tune the remaining 3D Gaussians after pruning. The table demonstrates that our method achieves better results across most of the pruning stages consistently than any other baselines. Similarly 18 for ACID dataset and 19 for RE10K dataset contains results for different feed-forward backbones where we finetune the remaining 3D Gaussians.

We’ve also included a detailed comparison of each scene with all baselines for a clearer understanding of our method’s performance. 9 ,10 shows the trends of PSNR, LPIPS and SSIM over different pruning percentages with and without finetuning on RE10K and ACID datasets with and without finetuning 3D Gaussians respectively with NoPoSplat backbone.

We can see from the plot that we consistently outperform all the baselines in all pruning scenarios 11, 12, 13, 14, 15 shows visual results on ACID dataset for pruning stages of \(\beta=0.4\) (\(40\%\)), \(\beta=0.5\) (\(50\%\)), \(\beta=0.6\) (\(60\%\)), \(\beta=0.7\) (\(70\%\)) and \(\beta=0.8\) (\(80\%\)) of 3D Gaussians where we take random pruning strategy on various feed-forward backbone. Similarly 16 17 , 18, 19, 20 shows results on RE10K dataset for different backbone with random pruning of 3D Gaussians.

21 presents an example scene evaluated at different 3D Gaussian counts on ACID and RE10K, where we use random pruning on different feed-forward backbone. Similarly, 22 shows the same scene compared against baselines incorporating the LightGaussian pruning method.

12 Cross domain evaluation↩︎

We also evaluate the cross domain performance of our model by training the model on one particular dataset and evaluating on the other. pecifically, we trained our model on the ACID dataset and evaluated on the DTU dataset; similarly, we also trained on the RE10K dataset and evaluated on the DTU dataset. 20 shows results where we train the baseline on ACID dataset and evaluate on DTU dataset whereas 21 is trained on RE10K dataset and evaluated on DTU dataset. The baseline that we use is specifically built on NoPoSplat where we use LightGaussian pruning as one baseline and random pruning as another. We use only NoPoSplat as pixelSplat and MVSplat are not trained on large scale diverse dataset as MASt3R used in NoPoSplat. The 23 shows that our method achieves superior results compared to baselines for the same reason as we discussed in main paper.

13 Failure Case↩︎

DWT limitation examples: As shown in 24, due to light-induced high frequencies and specular highlights, DWT provides a misleading signal by assigning high energy to Gaussians concentrated in a small sub-region within the marked flat SuperCluster. This results in noticeable degradation at high \(\beta\), whereas our method preserves quality in the remaining region. Addressing this limitation is left for future work.

14 Additional Ablation↩︎

14.1 Choice of k in K-Means Clustering↩︎

Our scene partitioning relies on k-means clustering to form SuperClusters. An ablation study on the ACID dataset (22) reveals that increasing k generally improves reconstruction quality. This is because a higher k yields finer-grained SuperClusters, enabling more precise identification of low-texture regions and, consequently, more judicious placement of larger 3D Gaussians for an optimized trade-off between sparsity and visual fidelity. However, increasing \(k\) beyond \(300\) yields only marginal improvements in quality while significantly increasing training time. 23 further reports the contribution of each component (GS Head training, contextual mask, structured pruning) on ACID at \(\beta=0.4\) and \(\beta=0.9\).

14.2 Choice of \(\gamma\) values and percentage of retained 3D Gaussians↩︎

24 presents the results of our method when retaining various percentages of 3D Gaussians within low-texture SuperClusters. We observe that retaining \(\gamma=0.10\) (\(10\%\)) of 3D Gaussians yields the best performance, with this optimal balance becoming particularly evident at high pruning levels (e.g., \(13k\) (\(\beta=0.9\)) 3D Gaussians). This is because aggressive pruning necessitates a precise balance of sparse, representative Gaussians that can effectively fill vacated regions without creating excessive overlap or redundancy. We notice that with increasing \(\gamma\) of primitives in a low-texture SuperCluster, we see a drop in the novel view metric. This is because increasing \(\gamma\) leads to redundant Gaussians in the same region, which leads to suboptimal quality. This insight is also exploited in previous works such as LightGaussian, MiniSplatting, GenerativeDensification, EAGLES, etc.

14.3 Pruning by choosing SuperClusters randomly:↩︎

Another ablation we perform is where we remove the texture aware pruning and choose SuperClusters randomly instead of taking texture into account. This way we are pruning the 3D Gaussians within a SuperCluster but the SuperCluster is chosen randomly. The results for the ablation are presented in 25 and 25 for RE10K dataset and in 26 for ACID dataset. The table shows that texture aware selection of SuperCluster is providing better results in most of the cases compared to random selection. 26 shows results for artifacts observed in case we use random SuperCluster for pruning rather than texture based method.

3pt

14.4 Additional Ablation on RE10K↩︎

We include extended ablations on RE10K across both moderate and aggressive pruning levels in 27 and 27. These results confirm that each module - Gaussian Adaptation, Masking, and Structured Pruning plays a distinct and essential role.

Key insights from RE10K ablations:

• Without Gaussian Adaptation: feed forward representation collapses under strong pruning (\(13K\)), showing that adapting the remaining Gaussians is critical for maintaining scene coverage.

• Without Mask: performs notably worse than the full model, demonstrating the importance of contextual, texture-aware selection rather than uniform or naïve pruning.

• AdaptiveSplat: consistently outperforms all ablations at high compression, validating the necessity of jointly learning masks and adaptive Gaussian updates.

References↩︎