Geometry-Aware Contrastive Feature Matching

• Geometry-Aware Contrastive Feature Matching (GCFM) is a method that integrates multi-modal features to enforce structured style transfer in 3D Gaussian Splatting.
• It employs a contrastive loss that pulls similar style features closer and pushes dissimilar ones apart, aligning appearance with geometric structure.
• GCFM operates within a decoupled optimization framework alternating between color and geometry updates to enhance stability and fidelity in stylized 3D scenes.

Geometry-Aware Contrastive Feature Matching (GCFM) is a geometry-sensitive feature matching objective introduced as the central guidance mechanism in a geometry-aware style transfer framework for 3D Gaussian Splatting (3DGS). Its purpose is to align a rendered view of a content scene with a style reference using multi-modal features built from RGB, depth, and edge cues, while using a contrastive formulation that pulls rendered features toward similar style features and pushes them away from dissimilar ones. Within that framework, GCFM is applied during both color-update and geometry-update phases of a decoupled optimization scheme, so that appearance stylization and structural deformation are supervised by the same matching principle (Bang et al., 23 Jun 2026).

1. Conceptual definition and problem setting

GCFM was introduced to address a specific limitation in 3DGS stylization: many earlier methods are described as appearance-driven, in the sense that they mostly alter Gaussian colors and leave geometry fixed or only weakly constrained. The consequence, as stated in the source work, is that such methods can preserve visual safety while failing to express the geometric character of a style image. The same work emphasizes that 3DGS output is especially sensitive to geometry changes, so naïvely optimizing color and geometry together can cause instability, including distorted geometry, over-stylized color, and scene-structure collapse (Bang et al., 23 Jun 2026).

In that setting, GCFM serves as a geometry-sensitive matching loss rather than a standalone matcher. It is designed to move beyond matching only RGB features by fusing color features for appearance, depth features for coarse spatial structure, and edge features for boundary and contour alignment. The contrastive form is motivated by the claim that positive correspondences alone are too weak for expressive 3D style transfer: the model should also push apart dissimilar style structures to obtain more discriminative geometry-aware representations. This suggests that GCFM is best understood not as a generic image-to-image contrastive loss, but as a structural alignment mechanism specialized to the instability of geometry manipulation in 3DGS (Bang et al., 23 Jun 2026).

A common misconception is to reduce GCFM to “style matching with extra modalities.” In the source formulation, its role is narrower and more technical: it is the supervisory signal that makes style transfer structurally expressive while remaining coupled to a controlled optimization schedule. The same discussion also makes clear that GCFM is geometry-aware rather than unconstrained geometry synthesis, because the strongest geometric stylization is still deliberately controlled to avoid scene collapse (Bang et al., 23 Jun 2026).

2. Contrastive formulation and matching rule

The full GCFM formulation is defined over rendered and style feature maps. For each spatial location $\mathbf{x}=(u,v)$ in the rendered feature map $\mathbf{F}^{\mathcal{R}}$, the method identifies a positive style location by nearest-neighbor search and a negative style location by farthest-neighbor search under cosine distance. The distance is defined as

$D_{\cos}(\cdot,\cdot) = 1 - \cos(\cdot,\cdot).$

The positive and negative matches are then

$$\mathbf{y}^{+}(\mathbf{x}) = \underset{\mathbf{y}}{\arg\min}\; D_{\cos}\big(\mathbf{F}^{\mathcal{R}}(\mathbf{x}), \mathbf{F}^{\mathcal{S}}(\mathbf{y})\big),$$

$$\mathbf{y}^{-}(\mathbf{x}) = \underset{\mathbf{y}}{\arg\max}\; D_{\cos}\big(\mathbf{F}^{\mathcal{R}}(\mathbf{x}), \mathbf{F}^{\mathcal{S}}(\mathbf{y})\big).$$

The resulting loss averages a contrastive term over all spatial positions in the feature map. In the authors’ formulation, the objective contains an explicit constant term $2+$, followed by the positive distance and the negative distance with opposite signs:

$$\mathcal{L}_{\mathrm{GC}} = \frac{1}{|\Omega_{\mathbf{F}}|} \sum_{\mathbf{x}\in\Omega_{\mathbf{F}}} \Big( 2 + D_{\cos}\big(\mathbf{F}^{\mathcal{R}}(\mathbf{x}), \mathbf{F}^{\mathcal{S}}(\mathbf{y}^{+}(\mathbf{x}))\big) - D_{\cos}\big(\mathbf{F}^{\mathcal{R}}(\mathbf{x}), \mathbf{F}^{\mathcal{S}}(\mathbf{y}^{-}(\mathbf{x}))\big) \Big).$$

The interpretation given in the source is direct. The first distance term pulls the rendered anchor feature toward a positive style match, the second pushes it away from a negative style match, and the constant “$2+$” is part of the exact objective as written. Positive and negative pairing are defined purely by feature distance in the joint embedding space, not by explicit semantic labels. A plausible implication is that GCFM relies on the quality of the fused feature space more than on any external correspondence annotations (Bang et al., 23 Jun 2026).

3. Multi-modal feature construction

The geometry-aware character of GCFM comes from how the feature maps are built. The rendered image and the style image are each encoded into a joint feature representation by channel-wise concatenation of color, depth, and edge features:

$$\mathbf{F}^{s} = \big[ \mathbf{F}_{\mathrm{color}^{s}}, \mathbf{F}_{\mathrm{depth}^{s}}, \mathbf{F}_{\mathrm{edge}^{s}} \big], \qquad s \in \{\mathcal{R}, \mathcal{S}\}.$$

The color branch uses a pretrained VGG network to obtain deep color-related features. These are described as capturing appearance, texture, and semantic similarity. Depth cues differ by domain: for the rendered image, depth comes directly from the 3DGS renderer, while for the style image it is estimated using Depth Anything V2. The source text notes that depth provides coarse geometric layout and spatial consistency, but also that depth alone can be noisy or incomplete (Bang et al., 23 Jun 2026).

Edge cues are added to strengthen boundary-level structure. For the rendered image, edges are extracted from the rendered appearance. For the style image, edges are extracted from the style image depth map rather than from the style image in the color domain, because stylized textures and fine patterns can make color-domain edges noisy. Canny edge detection is applied to these sources. The three modalities are concatenated into one joint representation, and matching is performed on that fused representation rather than on independently matched modalities, because independent matching can create inconsistencies between appearance, shape, and contours (Bang et al., 23 Jun 2026).

This design establishes an important distinction between GCFM and purely appearance-driven feature matching. GCFM does not treat geometry as a regularizer added after feature alignment; geometry is part of the representation that is being aligned. The source ablation further states that using only one modality leads to oversimplified geometric structures, whereas the full multi-modal version yields sharper contours and more faithful geometry, indicating that depth and edge cues are complementary rather than redundant (Bang et al., 23 Jun 2026).

4. Function within decoupled 3DGS optimization

GCFM operates inside a decoupled optimization scheme that alternates between updating color parameters and geometry parameters of Gaussian primitives. The parameters are decomposed as color parameters $\mathbf{\Theta} = \{\mathbf{\theta}_n\}$ and geometry parameters $\mathbf{F}^{\mathcal{R}}$0. The alternating updates are formalized as

$\mathbf{F}^{\mathcal{R}}$1

$\mathbf{F}^{\mathcal{R}}$2

Within this alternating loop, GCFM appears in both phases through the shared total objective

$\mathbf{F}^{\mathcal{R}}$3

with $\mathbf{F}^{\mathcal{R}}$4 during color optimization and $\mathbf{F}^{\mathcal{R}}$5 during geometry optimization. The reported loss weight for GCFM in the main experiments is $\mathbf{F}^{\mathcal{R}}$6 (Bang et al., 23 Jun 2026).

The operational pipeline is explicit. A view is first rendered from the current 3DGS parameters. For both the rendered view and the style image, VGG-based color features, depth-based features, and edge features are extracted and concatenated into a joint feature map. For every spatial location in the rendered feature map, the nearest style feature defines the positive and the farthest style feature defines the negative. The contrastive loss is then computed over all spatial positions. During color optimization, the loss includes $\mathbf{F}^{\mathcal{R}}$7, content loss, and TV loss. During geometry optimization, the same terms are used together with geometry regularization (Bang et al., 23 Jun 2026).

The significance of this shared use across both phases is central to the method’s design. In the authors’ description, GCFM makes the color phase learn style-consistent appearance features and the geometry phase learn style-consistent structural deformation. This suggests that the framework does not separate “appearance transfer” and “geometry transfer” into unrelated supervisory channels; instead, both are constrained by a single geometry-aware matching objective (Bang et al., 23 Jun 2026).

5. Empirical behavior, ablations, and practical constraints

The empirical discussion attributes several qualitative and quantitative effects to GCFM. In the ablation study, removing the contrastive term weakens feature alignment and causes the depth maps to lose structural distinctiveness. Single-modal variants yield oversimplified geometric structures, while the full RGB-depth-edge formulation produces sharper contours and more faithful geometry. Qualitatively, the method attributes sharper contours, stronger structural deformation, and more expressive geometric stylization to GCFM, with stylized scenes exhibiting style-specific structural patterns rather than merely recolored surfaces (Bang et al., 23 Jun 2026).

The source also reports a hyperparameter sensitivity study for the GCFM weight. Performance is described as stable across a broad range of $\mathbf{F}^{\mathcal{R}}$8, but setting $\mathbf{F}^{\mathcal{R}}$9 substantially degrades results. The reported SIFID values are summarized below (Bang et al., 23 Jun 2026).

$D_{\cos}(\cdot,\cdot) = 1 - \cos(\cdot,\cdot).$0	SIFID
0.0	1.5718
1.0	1.1683
2.0	1.1736
5.0	1.1841
10.0	1.1878

The full method, which combines GCFM with decoupled optimization, is reported to achieve the best overall quantitative performance, including the best SIFID among compared methods, strong multi-view consistency, and favorable user-study results. At the same time, the discussion implies several practical constraints. GCFM depends on reliable depth estimation for the style image, so errors in that stage can affect structural guidance. Edge extraction from depth is a heuristic adopted because color-domain edges can be noisy in stylized images. The method still requires a carefully balanced optimization schedule and geometry regularization to prevent instability; GCFM alone does not eliminate the difficulty of geometry updates. These points are important because they delimit what “geometry-aware” means in this context: the loss improves structural control, but it is not presented as sufficient by itself for arbitrary geometry manipulation (Bang et al., 23 Jun 2026).

6. Position within geometry-aware feature matching research

GCFM belongs to a broader line of work that attempts to make feature alignment sensitive to geometry rather than to appearance alone, but its specific formulation is tied to 3DGS style transfer rather than to classical correspondence estimation. The surrounding literature illustrates several neighboring design patterns.

Work	Core mechanism	Relation to GCFM
AdaMatcher (Huang et al., 2022)	Adaptive many-to-one, one-to-many, and one-to-one patch assignment	Shares the goal of correcting geometry-inconsistent matching under scale and viewpoint change
“All Graphs Lead to Rome” (Phillips et al., 2019)	GCN embedding with cycle consistency and epipolar geometric loss	A conceptual precursor: geometry-aware, but not contrastive in the modern positive-vs-negative sense
GECO (Hartwig et al., 1 Aug 2025)	OT-based soft assignment with dustbin and unbalanced OT	A geometry-aware feature learner emphasizing occlusions, partial assignment, and efficient inference
GAGeo (Wang et al., 29 Jun 2026)	Contrastive alignment of mask-pooled object features with satellite as anchor	Extends geometry-aware contrastive alignment into cross-view geo-localization with 3D priors and pose prediction

AdaMatcher is relevant because it identifies a different form of geometry inconsistency: mutual nearest neighbor or one-to-one patch assignment becomes invalid under large viewpoint or scale changes, since multiple patches in one image may legitimately map to the same patch in the other. Its solution is adaptive assignment, allowing many-to-one, one-to-many, and one-to-one correspondences. Although its domain is detector-free local feature matching rather than stylization, it shares with GCFM the principle that feature matching should reflect projective structure instead of forcing rigid appearance-based correspondence rules (Huang et al., 2022).

“All Graphs Lead to Rome” is described in the source material as close in spirit to geometry-aware contrastive feature matching, even though it predates the GCFM framing and is not explicitly contrastive in the modern metric-learning sense. It formulates multi-image matching as graph embedding, imposes cycle consistency through low-rank factorization, and adds geometric consistency through an epipolar loss during training. The connection is conceptual: both approaches use geometry to regularize representation learning, but the earlier method does so through graph reconstruction and cycle consistency rather than nearest-positive versus farthest-negative contrastive matching (Phillips et al., 2019).

GECO provides another neighboring formulation. It is a feature-learning method aimed at making dense descriptors geometry-aware rather than merely semantically similar, and it replaces hard argmax assignment with OT-based soft assignment, including a dustbin bin and unbalanced OT to model occlusions and disocclusions. In contrast to GCFM, GECO is explicitly framed as a representation-learning problem rather than as a matching loss inside a stylization pipeline. A plausible implication is that GCFM and GECO occupy complementary positions: one supplies geometry-aware supervision for 3D scene stylization, while the other learns reusable geometry-aware descriptors for downstream correspondence tasks (Hartwig et al., 1 Aug 2025).

GAGeo extends the contrastive pattern into cross-view object geo-localization. Its contrastive loss aligns mask-pooled object features across views and uses the satellite image as a universal anchor, while a permutation-equivariant 3D foundation-model backbone and explicit pose supervision constrain the learned space geometrically. Relative to GCFM, this shifts the matched entity from pixelwise fused style-transfer features to object-centric cross-view embeddings. The shared theme is that contrastive feature matching becomes geometry-aware only when the contrastive space is shaped by structural cues, view interaction, or 3D priors rather than by appearance similarity alone (Wang et al., 29 Jun 2026).

Taken together, these works suggest a broader research movement from appearance-dominant feature alignment toward objectives that encode structural validity, whether through adaptive assignment, cycle consistency, optimal transport, or contrastive cross-view alignment. Within that landscape, GCFM is specifically the mechanism that enables structurally expressive and optimization-stable 3DGS style transfer by coupling RGB, depth, and edge fusion to a nearest-neighbor versus farthest-neighbor contrastive objective (Bang et al., 23 Jun 2026).