💻 computer science

ConFi-GS Confidence-Guided High-Frequency Injection for 3D Gaussian Splatting Super-Resolution

This paper proposes ConFi-GS, a reliability-aware frequency modeling framework that enhances 3D Gaussian Splatting super-resolution by distinguishing between regions needing detail and those with reliable high-frequency content, thereby guiding a unified optimization scheme to inject consistent, high-fidelity textures while suppressing view-inconsistent artifacts.

Original authors: Jiaxiang Li, Zongtan Zhou, Zhen Tan, Yadong Liu, Dewen Hu

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiaxiang Li, Zongtan Zhou, Zhen Tan, Yadong Liu, Dewen Hu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to rebuild a detailed 3D model of a city, but you only have a few blurry, low-resolution photos of it. If you try to guess what the missing details look like, you might accidentally invent things that aren't there—like painting a fake window on a wall that's actually solid, or making a tree look sharp in one photo but blurry in another. This is the problem the paper "ConFi-GS" tries to solve for a technology called 3D Gaussian Splatting (a method for creating 3D scenes from photos).

Here is a simple breakdown of how their solution works, using everyday analogies:

The Core Problem: "Hunger" vs. "Trust"

When you try to fix a blurry 3D model, you face two different questions:

Where is the model hungry for detail? (Which parts look blurry and need help?)
Can we trust the new details we are about to add? (Is the "high-resolution" reference we are looking at actually accurate, or is it just a hallucination?)

Existing methods often just say, "This area looks blurry, so let's paste high-quality details there immediately." The problem is that sometimes the "high-quality" details come from a single photo that has weird artifacts or inconsistencies. If you paste those in, your 3D model becomes sharp in one view but looks broken or inconsistent when you look at it from another angle.

The Solution: A Smart "Detail Injection" System

The authors propose a system called ConFi-GS that acts like a very strict, smart editor. Instead of blindly pasting details, it uses a three-step filter to decide what gets added to the 3D model.

1. The "Hunger Map" (Geometry-Guided Demand)

First, the system looks at the 3D model and asks: "Where are we actually missing information?"

Analogy: Imagine a construction crew looking at a building. They check the blueprints and see that the north wall has been photographed from many angles, so they know exactly what it looks like. However, the south wall has only been seen from a distance and looks fuzzy. The system marks the south wall as "Hunger Zone"—a place that genuinely needs more detail.

2. The "Trust Filter" (Frequency-Aware Reliability)

Next, even if a spot is "hungry," the system checks if the new details it's about to add are trustworthy. It looks at the high-resolution reference images and asks three questions:

Does it make structural sense? (Is the detail aligned with real edges, like a door frame, or is it just random noise?)
Is it actually missing? (Does the current 3D model actually lack this specific high-frequency pattern, or is it already there?)
Is it consistent? (If we look at this detail from a slightly different angle, does it still look the same, or does it shift and warp?)
Analogy: Imagine a detective trying to solve a crime. Just because a witness (the high-res photo) says they saw a red car doesn't mean they should write it down immediately. The detective checks: Is the witness reliable? Does the story match other evidence? Is the red car consistent with the timeline? If the witness is shaky or the story doesn't add up, the detective ignores that detail.

3. The "Injection Map" (The Final Decision)

The system combines the "Hunger Map" and the "Trust Filter" to create a final Detail Injection Map.

Rule: We only add high-quality details to a spot if BOTH conditions are met: The spot is hungry for detail AND the new detail is proven to be reliable.
Result: This prevents the 3D model from "hallucinating" fake textures or becoming inconsistent when viewed from different angles.

How the System Learns (The Training Process)

Once the system knows where to add details, it uses a smart training strategy to actually put them there:

Coarse-to-Fine Learning: Instead of trying to learn every tiny detail on day one, the system starts by learning the big shapes and colors (the "coarse" stuff). Only after the structure is stable does it start adding the fine, sharp details (the "high-frequency" stuff).
- Analogy: Think of painting a portrait. You first sketch the outline and block in the basic colors. You don't try to paint the individual eyelashes until the face shape is perfect. If you try to paint eyelashes too early, you might mess up the whole face.
Smart Expansion: If the system finds a reliable detail that the current 3D model is too "coarse" to capture, it adds more "building blocks" (Gaussians) specifically to that area.
- Analogy: If you are trying to describe a bumpy road, and your current description is too smooth, you don't just smooth it out more. You add more specific notes about the bumps, but only where the bumps actually exist and are verified.

The Result

By using this "Reliability-Aware" approach, the paper shows that they can turn blurry, low-resolution photos into sharp, high-quality 3D models. Crucially, these models look sharp from every angle and don't contain the weird, flickering artifacts that happen when you try to force details into a model that isn't ready for them.

In short: They built a system that knows exactly where to look for missing details and how to verify that those details are real before adding them, ensuring the final 3D world is both sharp and stable.

Technical Summary: ConFi-GS: Confidence-Guided High-Frequency Injection for 3D Gaussian Splatting Super-Resolution

Problem Statement
Reconstructing high-quality 3D scenes from low-resolution multi-view images using 3D Gaussian Splatting (3DGS) is hindered by insufficient high-frequency observations, leading to blurred textures, weak boundaries, and view-inconsistent details. While existing approaches attempt to address this by applying super-resolution (SR) guidance uniformly or localizing enhancement based on geometric sampling, they fail to distinguish between two critical questions: where additional detail is needed, and whether the corresponding candidate high-frequency content is reliable enough to be internalized into a multi-view consistent 3D representation. Directly enforcing super-resolved details without reliability assessment can introduce view-specific artifacts, structurally unsupported sharpness, and hallucinated textures that degrade the 3D representation's consistency.

Methodology
The authors propose ConFi-GS, a reliability-aware frequency modeling framework that reformulates low-resolution 3DGS reconstruction as a problem of selective detail injection. The framework operates through four coordinated components:

Geometry-Guided Detail Demand Estimation:
The method first pretrains a base 3DGS model on low-resolution inputs to estimate a "detail-demand prior." It analyzes the cross-view sampling ratio of each Gaussian primitive. If a Gaussian is sampled at a similar, insufficient scale across all views (low sampling ratio), it is flagged as under-detailed. This generates a view-specific detail-demand map ( $D_t$ ) that highlights regions likely lacking recoverable detail, independent of the reliability of the candidate details.
Frequency-Aware Reliability Assessment:
To determine if candidate high-frequency details from a super-resolved reference are trustworthy, the framework computes a reliability map ( $C^t_{rel}$ ) based on three cues:
- Structural Support ( $E_t$ ): Measures alignment with local edges in the super-resolved reference to ensure details are not structurally unsupported.
- Unresolved Evidence ( $G_t$ ): Quantifies the high-frequency discrepancy between the current rendering and the super-resolved reference using a Fourier-domain high-pass operator, identifying spectral content that is missing but potentially recoverable.
- Cross-View Consistency ( $X_t$ ): Penalizes details that are unstable when reprojected across neighboring views, filtering out view-specific artifacts.
  The final detail-injection map ( $M_t$ ) is the element-wise product of the demand map and the reliability map, ensuring supervision is applied only where detail is both needed and reliable.
Selective Progressive Frequency Regularization:
Instead of enforcing full-band frequency matching immediately, the method employs a coarse-to-fine schedule. It applies patch-level amplitude-spectrum loss ( $L_{amp}$ ) and phase-spectrum loss ( $L_{ph}$ ) only to regions with high weights in $M_t$ . The amplitude loss is introduced early to stabilize coarse appearance, while the more sensitive phase loss is activated later once the geometry is stable. This prevents the model from fitting unstable pseudo-high-frequency signals during early optimization.
Reliability-Aware Gaussian Densification:
To ensure the representation has the capacity to absorb reliable details, the authors introduce a densification strategy guided by a joint score ( $S_i$ ) combining geometric demand ( $U_i$ ), reliability-weighted spectral residuals ( $F_i$ ), and local structural complexity ( $H_i$ ). This directs the splitting and cloning of Gaussians specifically to regions where unresolved, reliable details exist, rather than relying solely on generic image-space gradients.

Key Contributions

Reliability-Aware Problem Formulation: The paper explicitly separates "detail demand" from "detail reliability," addressing the fundamental ambiguity in low-resolution 3D reconstruction where missing detail does not guarantee that super-resolved candidates are trustworthy.
Dual-Stage Estimation: It introduces a geometry-guided demand estimator followed by a frequency-aware reliability assessment that evaluates structural support, spectral resolution, and cross-view stability to produce a precise detail-injection map.
Unified Optimization Scheme: The framework integrates selective progressive frequency regularization and reliability-aware Gaussian densification, enabling the gradual and structural internalization of reliable high-frequency details into the 3D representation.

Experimental Results
The method was evaluated on standard benchmarks including Tanks & Temples, Deep Blending, and Mip-NeRF 360 under 2× and 4× super-resolution settings.

Quantitative Performance: ConFi-GS achieved state-of-the-art performance across multiple metrics (PSNR, SSIM, LPIPS, FID, CMMD, DreamSim) compared to baselines such as standard 3DGS (LR), Mip-Splatting, SRGS (uniform SR), and SplatSuRe (geometry-guided SR). For instance, on Tanks & Temples (4×), it achieved a PSNR of 24.08 and LPIPS of 0.270, outperforming SplatSuRe (23.84 PSNR, 0.273 LPIPS).
Qualitative Improvements: Visual comparisons demonstrated sharper reconstructions around thin boundaries and textured areas while effectively suppressing high-frequency artifacts and view-inconsistent details that plagued other methods.
Ablation Studies: Removing any component (reliability assessment, progressive regularization, or reliability-aware densification) resulted in significant performance degradation, confirming the complementary necessity of each module. Specifically, removing reliability assessment led to the absorption of view-inconsistent details, while removing progressive scheduling destabilized early optimization.

Significance and Claims
The paper claims that ConFi-GS shifts the paradigm of low-resolution 3DGS enhancement from "direct detail amplification" to "reliability-aware detail internalization." By explicitly filtering out unreliable high-frequency signals before they are internalized into the 3D representation, the method achieves a balance between high-fidelity texture recovery and multi-view structural consistency. The authors emphasize that their approach does not merely imitate super-resolved image textures but progressively integrates trustworthy details into a stable 3D scene representation. While acknowledging limitations such as dependence on external SR quality and additional training time, the work establishes a framework for evidence-aware detail integration that addresses the core ambiguity of detail demand versus reliability in 3D reconstruction.