SRSplat: Feed-Forward Super-Resolution Gaussian Splatting from Sparse Multi-View Images
Extracted source text · arXiv 2511.12040
%File: formatting-instructions-latex-2026.tex
%release 2026.0
\documentclass[letterpaper]{article} % DO NOT CHANGE THIS
\usepackage{aaai2026} % DO NOT CHANGE THIS
\usepackage{times} % DO NOT CHANGE THIS
\usepackage{helvet} % DO NOT CHANGE THIS
\usepackage{courier} % DO NOT CHANGE THIS
\usepackage[hyphens]{url} % DO NOT CHANGE THIS
\usepackage{graphicx} % DO NOT CHANGE THIS
\urlstyle{rm} % DO NOT CHANGE THIS
\def\UrlFont{\rm} % DO NOT CHANGE THIS
\usepackage{natbib} % DO NOT CHANGE THIS AND DO NOT ADD ANY OPTIONS TO IT
\usepackage{caption} % DO NOT CHANGE THIS AND DO NOT ADD ANY OPTIONS TO IT
\frenchspacing % DO NOT CHANGE THIS
\setlength{\pdfpagewidth}{8.5in} % DO NOT CHANGE THIS
\setlength{\pdfpageheight}{11in} % DO NOT CHANGE THIS
%
% These are recommended to typeset algorithms but not required. See the subsubsection on algorithms. Remove them if you don't have algorithms in your paper.
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{mathtools}
\usepackage{ragged2e}
\usepackage{caption}
\usepackage{subcaption}
\usepackage{booktabs}
\usepackage{multirow}
\usepackage{makecell}
\usepackage{colortbl}
\usepackage{float}
\usepackage{array}
\usepackage{diagbox}
\usepackage{tabularx} % 自动列宽
\usepackage{siunitx} % 数字对齐
\newcommand\ChangeRT[1]{\noalign{\hrule height #1}}
\sisetup{
detect-mode,
table-number-alignment = center,
round-mode = places,
round-precision = 2,
table-format = 2.2
}
%
% These are are recommended to typeset listings but not required. See the subsubsection on listing. Remove this block if you don't have listings in your paper.
\usepackage{newfloat}
\usepackage{listings}
\DeclareCaptionStyle{ruled}{labelfont=normalfont,labelsep=colon,strut=off} % DO NOT CHANGE THIS
\lstset{%
basicstyle={\footnotesize\ttfamily},% footnotesize acceptable for monospace
numbers=left,numberstyle=\footnotesize,xleftmargin=2em,% show line numbers, remove this entire line if you don't want the numbers.
aboveskip=0pt,belowskip=0pt,%
showstringspaces=false,tabsize=2,breaklines=true}
\floatstyle{ruled}
\newfloat{listing}{tb}{lst}{}
\floatname{listing}{Listing}
%
% Keep the \pdfinfo as shown here. There's no need
% for you to add the /Title and /Author tags.
\pdfinfo{
/TemplateVersion (2026.1)
}
% DISALLOWED PACKAGES
% \usepackage{authblk} -- This package is specifically forbidden
% \usepackage{balance} -- This package is specifically forbidden
% \usepackage{color (if used in text)
% \usepackage{CJK} -- This package is specifically forbidden
% \usepackage{float} -- This package is specifically forbidden
% \usepackage{flushend} -- This package is specifically forbidden
% \usepackage{fontenc} -- This package is specifically forbidden
% \usepackage{fullpage} -- This package is specifically forbidden
% \usepackage{geometry} -- This package is specifically forbidden
% \usepackage{grffile} -- This package is specifically forbidden
% \usepackage{hyperref} -- This package is specifically forbidden
% \usepackage{navigator} -- This package is specifically forbidden
% (or any other package that embeds links such as navigator or hyperref)
% \indentfirst} -- This package is specifically forbidden
% \layout} -- This package is specifically forbidden
% \multicol} -- This package is specifically forbidden
% \nameref} -- This package is specifically forbidden
% \usepackage{savetrees} -- This package is specifically forbidden
% \usepackage{setspace} -- This package is specifically forbidden
% \usepackage{stfloats} -- This package is specifically forbidden
% \usepackage{tabu} -- This package is specifically forbidden
% \usepackage{titlesec} -- This package is specifically forbidden
% \usepackage{tocbibind} -- This package is specifically forbidden
% \usepackage{ulem} -- This package is specifically forbidden
% \usepackage{wrapfig} -- This package is specifically forbidden
% DISALLOWED COMMANDS
% \nocopyright -- Your paper will not be published if you use this command
% \addtolength -- This command may not be used
% \balance -- This command may not be used
% \baselinestretch -- Your paper will not be published if you use this command
% \clearpage -- No page breaks of any kind may be used for the final version of your paper
% \columnsep -- This command may not be used
% \newpage -- No page breaks of any kind may be used for the final version of your paper
% \pagebreak -- No page breaks of any kind may be used for the final version of your paperr
% \pagestyle -- This command may not be used
% \tiny -- This is not an acceptable font size.
% \vspace{- -- No negative value may be used in proximity of a caption, figure, table, section, subsection, subsubsection, or reference
% \vskip{- -- No negative value may be used to alter spacing above or below a caption, figure, table, section, subsection, subsubsection, or reference
\setcounter{secnumdepth}{2} %May be changed to 1 or 2 if section numbers are desired.
% The file aaai2026.sty is the style file for AAAI Press
% proceedings, working notes, and technical reports.
%
% Title
% Your title must be in mixed case, not sentence case.
% That means all verbs (including short verbs like be, is, using,and go),
% nouns, adverbs, adjectives should be capitalized, including both words in hyphenated terms, while
% articles, conjunctions, and prepositions are lower case unless they
% directly follow a colon or long dash
\title{SRSplat: Feed-Forward Super-Resolution Gaussian Splatting from \\ Sparse Multi-View Images}
\author{
%Authors
% All authors must be in the same font size and format.
% Written by AAAI Press Staff\textsuperscript{\rm 1}\thanks{With help from the AAAI Publications Committee.}\\
% AAAI Style Contributions by Pater Patel Schneider,
% Sunil Issar,\\
% J. Scott Penberthy,
% George Ferguson,
% Hans Guesgen,
% Francisco Cruz\equalcontrib,
% Marc Pujol-Gonzalez\equalcontrib
Xinyuan Hu\textsuperscript{\rm 1} \equalcontrib,
Changyue Shi\textsuperscript{\rm 1} \textsuperscript{\rm 2} \equalcontrib,
Chuxiao Yang\textsuperscript{\rm 1},
Minghao Chen\textsuperscript{\rm 1},
Jiajun Ding\textsuperscript{\rm 1} \thanks{Corresponding author.}, \\
Tao Wei\textsuperscript{\rm 2} \textsuperscript{\rm 3},
Chen Wei\textsuperscript{\rm 3},
Zhou Yu\textsuperscript{\rm 1},
Min Tan\textsuperscript{\rm 1}
}
\affiliations{
%Afiliations
\textsuperscript{\rm 1}School of Computer Science and Technology, Hangzhou Dianzi University, \\
\textsuperscript{\rm 2}School of AI for Science, Peking University, \\
\textsuperscript{\rm 3}Li Auto Inc. \\
% If you have multiple authors and multiple affiliations
% use superscripts in text and roman font to identify them.
% For example,
% Sunil Issar\textsuperscript{\rm 2},
% J. Scott Penberthy\textsuperscript{\rm 3},
% George Ferguson\textsuperscript{\rm 4},
% Hans Guesgen\textsuperscript{\rm 5}
% Note that the comma should be placed after the superscript
% 1101 Pennsylvania Ave, NW Suite 300\\
% Washington, DC 20004 USA\\
% email address must be in roman text type, not monospace or sans serif
% proceedings-questions@aaai.org
\{xinyuan, shicy, chuxiao\_yang, djj\}@hdu.edu.cn
%
% See more examples next
}
%Example, Single Author, ->> remove \iffalse,\fi and place them surrounding AAAI title to use it
\iffalse
\title{My Publication Title --- Single Author}
\author {
Author Name
}
\affiliations{
Affiliation\\
Affiliation Line 2\\
name@example.com
}
\fi
\iffalse
%Example, Multiple Authors, ->> remove \iffalse,\fi and place them surrounding AAAI title to use it
\title{My Publication Title --- Multiple Authors}
\author {
% Authors
First Author Name\textsuperscript{\rm 1,\rm 2},
Second Author Name\textsuperscript{\rm 2},
Third Author Name\textsuperscript{\rm 1}
}
\affiliations {
% Affiliations
\textsuperscript{\rm 1}Affiliation 1\\
\textsuperscript{\rm 2}Affiliation 2\\
firstAuthor@affiliation1.com, secondAuthor@affilation2.com, thirdAuthor@affiliation1.com
}
\fi
% REMOVE THIS: bibentry
% This is only needed to show inline citations in the guidelines document. You should not need it and can safely delete it.
\usepackage{bibentry}
% END REMOVE bibentry
\begin{document}
\maketitle
\begin{abstract}
Feed-forward 3D reconstruction from sparse, low-resolution (LR) images is a crucial capability for real-world applications, such as autonomous driving and embodied AI. However, existing methods often fail to recover fine texture details. This limitation stems from the inherent lack of high-frequency information in LR inputs. To address this, we propose \textbf{SRSplat}, a feed-forward framework that reconstructs high-resolution 3D scenes from only a few LR views. Our main insight is to compensate for the deficiency of texture information by jointly leveraging external high-quality reference images and internal texture cues. We first construct a scene-specific reference gallery, generated for each scene using Multimodal Large Language Models (MLLMs) and diffusion models. To integrate this external information, we introduce the \textit{Reference-Guided Feature Enhancement (RGFE)} module, which aligns and fuses features from the LR input images and their reference twin image. Subsequently, we train a decoder to predict the Gaussian primitives using the multi-view fused feature obtained from \textit{RGFE}. To further refine predicted Gaussian primitives, we introduce \textit{Texture-Aware Density Control (TADC)}, which adaptively adjusts Gaussian density based on the internal texture richness of the LR inputs. Extensive experiments demonstrate that our SRSplat outperforms existing methods on various datasets, including RealEstate10K, ACID, and DTU, and exhibits strong cross-dataset and cross-resolution generalization capabilities.
\end{abstract}
% Uncomment the following to link to your code, datasets, an extended version or similar.
% You must keep this block between (not within) the abstract and the main body of the paper.
\begin{links}
\link{Project Page}{https://xinyuanhu66.github.io/SRSplat/}
% \link{Datasets}{https://aaai.org/example/datasets}
% \link{Extended version}{https://aaai.org/example/extended-version}
\end{links}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{Fig_cmyk/firstpipline1113_cmyk.pdf}
% \vspace{-2em}
\caption{
In this work, we propose SRSplat, a novel feed-forward framework that reconstructs high-quality 3D scenes with only sparse and LR input views. SRSplat demonstrates superior performance and is capable of handling low-resolution, sparse-view inputs and real-time reconstruction, thereby offering greater functionality and practicality in realistic applications.}
\label{fig:dive}
\end{figure*}
\section{Introduction}
Reconstructing 3D scenes from 2D images is a fundamental task in computer vision and computer graphics~\cite{mildenhall2021nerf,chen2022tensorf,kerbl20233d, shi2025mmgs,shi2025realm,muller2022instant}. It plays a crucial role in various applications, such as embodied AI~\cite{huang2023visual} and autonomous driving~\cite{tian2025drivingforward}. However, in these real-world applications, traditional 3D reconstruction methods like NeRF~\cite{mildenhall2021nerf} or 3DGS~\cite{kerbl20233d} face several key challenges: 1) \textbf{\textit{Resolution Constraints:}} Due to hardware and sensor limitations, acquiring sufficiently high-resolution (HR) images for accurate 3D reconstruction is often impractical.
2) \textbf{\textit{Few-Shot Challenges:}} In practice, acquiring high-quality dense views is prohibitively costly and impractical.
3) \textbf{\textit{Real-Time Reconstruction:}} Fields like robotics require real-time 3D reconstruction, necessitating efficient feed-forward algorithms.
Our goal is to build a feed-forward 3D reconstruction framework that simultaneously tackles the three practical hurdles so that high-quality 3D scenes can be reconstructed from only limited LR images. The recent emergence of feed-forward 3DGS~\cite{charatan2024pixelsplat,chen2024mvsplat,zhang2025transplat,tang2024hisplat,chen2024mvsplat360,fei2024pixelgaussian} methods has revolutionized 3D scene reconstruction through their real-time reconstruction capabilities. These methods leverage feed-forward networks to make direct 3D Gaussian predictions, eliminating the need for per-scene optimization.
However, when provided with low-resolution (LR) input images, existing methods struggle to reconstruct high-quality scenes and often exhibit a loss of texture details. This limitation stems from the inherent lack of high-frequency texture information in LR images compared to their HR counterparts~\cite{2024srgs}.
To this end, we propose \textbf{SRSplat}, a novel feed-forward framework that reconstructs HR 3D scenes from sparse LR views. Our main insight is to compensate for the deficiency of texture information by jointly leveraging external high-quality reference images and internal texture cues.
Inspired by previous reference-based 2D image super-resolution (SR) methods~\cite{lu2021masa,sun2024coser,yang2020learning,bosiger2024mariner,jiang2021robust,cao2022reference}, we first construct a scene-specific \textit{Reference Gallery}. In this gallery, each reference image is generated as a twin image of its corresponding input scene. Specifically, given a scene, we first employ Multimodal Large Language Models (MLLMs)~\cite{achiam2023gpt} to generate a concise semantic description. The resulting description is then used to prompt a pre-trained diffusion model~\cite{labs2025flux1kontextflowmatching} to synthesize high-quality reference twins. Our novel reference twin generation technique synthesizes images that mirror the target scene’s characteristics but in a HR way, supplying more high-frequency cues for the subsequent process.
With the \textit{Reference Gallery}, we propose the \textit{Reference-Guided Feature Enhancement (RGFE)} module to integrate this external information. For the input images and their corresponding HR reference twins, we first use a shared CNN for multi-scale feature extraction. Then \textit{RGFE} performs coarse-to-fine correspondence matching for the extracted features. Finally, a fuse network is used to map the distribution of reference features to LR features, thereby transferring high-frequency information from the reference twins.
The resulting fused features are then fed to a decoder to predict the Gaussian primitives. However, since the proposed Gaussian prediction decoder predicts a single Gaussian primitive for each pixel, texture-rich regions are difficult to optimize. To further refine predicted Gaussian primitives, we propose \textit{Texture-Aware Density Control (TADC)}. Specifically, \textit{TADC} builds a learnable texture richness perceptron that perceives texture richness from LR inputs. Based on the internal texture richness, Gaussian primitives can adaptively adjust the density in the scene.
Our contributions can be summarized as follows:
\begin{itemize}
\item We propose SRSplat, a feed-forward framework that generates HR 3D scenes from sparse LR 2D images. To the best of our knowledge, this is the first work to solve this task in a feed-forward manner.
\item We construct a scene-specific reference twin gallery using MLLMs and diffusion priors. We propose the \textit{RGFE} and \textit{TADC} module to compensate for the deficiency of texture information by leveraging external reference image and internal texture richness map.
\item Experimental results on various public datasets demonstrate that SRSplat outperforms existing methods and exhibits strong cross-dataset and cross-resolution generalization capabilities.
\end{itemize}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{Fig_cmyk/allpipline0723_cmyk.pdf}
% \vspace{-15pt}
\caption{Framework of SRSplat. Our method takes LR images and their corresponding reference as inputs. The \textit{RGFE} module first extracts multi-scale features and effectively fuses these features. Upon decoding the Gaussian primitives, \textit{TADC} adjusts Gaussian density adaptively according to the richness of texture generated by a texture richness perceptron.}
% \vspace{-15pt}
\label{fig:pipeline}
\end{figure*}
\section{Related Works}
\noindent \textbf{Feed-Forward 3D Reconstruction.} Recently, 3DGS~\cite{kerbl20233d}-based feed-forward reconstruction has emerged as a key approach for efficient 3D scene representation and novel view synthesis. PixelSplat~\cite{charatan2024pixelsplat} introduces a polar line-based Transformer architecture to model cross-view correspondence and predict depth distribution. MVSplat~\cite{chen2024mvsplat} constructs cost volumes via plane sweeping to achieve enhanced geometric reconstruction accuracy. TranSplat~\cite{zhang2025transplat} adopts a Transformer-based architecture, combining monocular depth priors to refine the sparse-view reconstruction results. DepthSplat~\cite{xu2025depthsplat} presents a robust multi-view depth model by leveraging pre-trained monocular depth features~\cite{yang2024depth}, thereby enabling high-quality feed-forward reconstructions. Despite these advancements, existing methods often fail to capture high-frequency details from low-resolution inputs, leading to artifacts and reduced quality. To address these limitations, we propose SRSplat, a novel framework that enables super-resolution reconstruction with low-resolution inputs.
\noindent \textbf{Super-Resolution Novel View Synthesis.} Super-resolution novel view synthesis aims to reconstruct high-resolution 3D scenes from only low-resolution multi-view inputs. As a pioneer in the field, NeRF-SR~\cite{wang2021nerf-sr} leverages the sub-pixel constraint to optimize a HR scenes. SRGS~\cite{2024srgs} is the first framework to synthesize HR novel views based on 3DGS using the SwinIR model~\cite{liang2021swinir}, but suffers from multi-view inconsistencies. Recently, S2Gaussian~\cite{wan2025s2gaussian} employs a multi-stage training paradigm to reconstruct high-resolution 3D scenes solely from only sparse and low-resolution input views. However, per-scene optimization methods rely on iterative optimization to achieve the final 3D representation and exhibit poor generalization ability. In contrast, feed-forward inference methods reconstruct the entire scene in a single feed-forward pass, demonstrating strong potential in super-resolution novel-view synthesis.
\noindent \textbf{Reference-Based Super-Resolution.}
Reference‐based SR methods aim to enhance low‐resolution inputs by leveraging high‐frequency details from high‐quality reference images. MASA‐SR~\cite{lu2021masa} introduces attention mechanisms to adaptively fuse reference textures into LR feature representations. More recently, RefSR-NeRF~\cite{huang2023refsr} further extends this idea to 3D by integrating reference images as auxiliary information. Yet, in real-world scenarios, it is often impractical to assume that every scene can be paired with a corresponding reference image. CoSeR~\cite{sun2024coser} leverages prior knowledge from large-scale text-to-image diffusion models to synthesize high-quality reference images for super-resolution. This motivates the development of an automated reference-generation approach.
\begin{figure*}
\centering
\includegraphics[width=0.98\linewidth]{Fig_cmyk/refgpt0722_cmyk.pdf}
% \vspace{-30pt}
\caption{Pipeline of reference gallery generation. Given LR input images for each scene, the MLLM produces semantic descriptions. Subsequently, the diffusion model uses these descriptions to generate reference images tailored to the scene. % high-quality
}
% \vspace{-5pt}
\label{fig:refgpt}
\end{figure*}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Fig_cmyk/ref_gal0731_cmyk.pdf}
% \vspace{-10pt}
\caption{Reference gallery examples. Reference twin images share details similar to the LR images.
}
% \vspace{-10pt}
\label{fig:refgal}
\end{figure}
\section{Framework of SRSplat}
The overall framework of SRSplat is illustrated in Fig.~\ref{fig:pipeline}.
We first establish a reference gallery in Sec.~\ref {sec:ref_gallery}.
Then we employ \textit{RGFE} to fuse features derived from input and reference images in Sec.~\ref{sec:RGFE}. Based on the fused features, we train a decoder to predict Gaussian attributes in Sec.~\ref {depth estimation}. Finally, we use \textit{TADC} to adaptively adjust the density of Gaussian primitives according to texture intensity in Sec.~\ref {sec:TADC}.
\subsection{Reference Gallery Preparation}
\label{sec:ref_gallery}
Drawing inspiration from 2D reference-based SR methods~\cite{lu2021masa,bosiger2024mariner}, we aim to mitigate the high-frequency loss of LR input multi-view images with the information from external reference images. To achieve this, we first establish a reference gallery. Each image in this gallery serves as a twin of each 3D scene, mirroring its characteristic layouts. The process is illustrated in Fig.~\ref{fig:refgpt}.
We first leverage GPT-4o to generate semantic descriptions of input LR multi-view images.
For a set of LR input images (downsampled by a factor of $P$ from training set)
of a specific scene $\{\,I^\text{LR}_i\}_{i=1}^{N},
I_i \in \mathbb{R}^{\frac{H}{P} \times \frac{W}{P} \times 3}$,
we employ MLLM to directly process the LR image set and generate semantic descriptions $ \mathcal{P} = {\text{MLLM}}(\{\,I^\text{LR}_i\}_{i=1}^{N})$.
The generated descriptions $\mathcal{P}$ capture the overall layout of the scene, key objects and their relationships, providing clear semantic guidance for the subsequent process.
With the semantic descriptions, we utilize a 2D diffusion model~\cite{labs2025flux1kontextflowmatching} $\mathcal{D}$ to generate a HR reference twin $r = \mathcal{D}(\mathcal{P}), r \in \mathbb{R}^{H \times W \times 3}$. By leveraging strong priors of MLLM and diffusion model, the reference twin faithfully preserves scene's layout and remains semantically aligned with the original inputs (as demonstrated in Fig.~\ref{fig:refgal}).
\subsection{Reference-Guided Feature Enhancement}
\label{sec:RGFE}
\noindent \textbf{\textit{Multi-Scale Feature Extraction.}} For an LR image $I^\text{LR} \in\mathbb{R}^{\frac{H}{P}\times \frac{W}{P}\times 3}$, we employ a shared CNN encoder~\cite{xu2022gmflow,xu2023unifying} $\mathcal{E}$ to extract multiscale features from both the upsampled input images $I\bigr\uparrow\in\mathbb{R}^{H\times W\times 3}$ and its corresponding reference $r\in\mathbb{R}^{H\times W\times 3}$. In practice, we employ three levels, each halving the resolution of the previous one. The encoded outputs are
\begin{equation}
\label{eq:feature}
\left\{
\begin{aligned}
\{F_{l}^{\mathrm{I}}\} &= \mathcal{E}(I\bigr\uparrow),\\
\{F_{l}^{\mathrm{ref}}\} &= \mathcal{E}(r),
\end{aligned}
\right.\quad l = 1,2,3,
\end{equation}
where $F_{l}^{\mathrm{I}},\,F_{l}^{\mathrm{ref}}\in\mathbb{R}^{H_{l}\times W_{l}\times C^{{feature}}}$ denote the features at scale $l$ with resolution $H_{l}\times W_{l}$ and channel $C^{feature}$.
\noindent \textbf{\textit{Feature Matching and Fusion.}} Following previous researches~\cite{lu2021masa,bosiger2024mariner}, SRSplat matchs the features between input and reference features using cosine similarity. Specifically, it first performs coarse-to-fine matching, starting with a coarse grid using a stride, followed by dense matching within a fixed-size window around the initial correspondences. This procedure yields a mapping \(m\) from input feature indices to corresponding reference feature indices:
\begin{equation}
\label{eq:masa}
m_{I\to ref}^l : (x,y)\in{F}_{l}^{I} \;\longmapsto\; \bigl\{(u,v)\in{F}_{l}^{ref},\;s\in\mathbb{R}\bigr\},
\end{equation}
where $(u,v)$ is the best match coordinate in the reference for the input coordinate $(x,y)$ and $s$ is the corresponding match score. Next, we warp the reference features according to \(m_{I\to ref}^l\) and weight them by their match scores, yielding the warped reference features \(\{F_1^{\mathrm{ref}\to\mathrm{I}}, F_2^{\mathrm{ref}\to\mathrm{I}}, F_3^{\mathrm{ref}\to\mathrm{I}}\}\). Weighting the warped features by their matching scores reduces low-confidence correspondences. This enables the model to use the reference features only when they have a confident match.
Finally, we fuse the warped reference features \(F_{l}^{\mathrm{ref}\!\rightarrow\!\mathrm{I}}\) with the image features \(F_{l}^{\mathrm{I}}\) using a learnable fusion network $\mathcal{H}$:
\begin{equation}
\label{eq:enhanced}
F_{enhanced} = \mathcal{H}\bigl( \{F_l^I\},\,\{F_l^{ref\to I}\}\bigr),
\end{equation}
where $F_{enhanced}\in\mathbb{R}^{\frac{H}{4}\times \frac{W}{4}\times C^{enhanced}}$ represents the final enhanced feature. To further enhance information exchange across views, we feed the features into the Swin Transformer~\cite{liu2021swin}, which employs both cross-attention and self-attention mechanisms.
\subsection{Depth Estimation and Gaussian Prediction}
\label{depth estimation}
We utilize the cost volume matching in Multi-View Stereo (MVS)~\cite{yao2018mvsnet,cao2022mvsformer} for depth estimation. Specifically, we first construct a depth candidate \(\displaystyle d^i_{\mathrm{cand}}\in\mathbb{R}^{\frac{H}{4}\times \frac{W}{4}\times D}\)
using the plane-sweep stereo approach~\cite{yao2018mvsnet}. Then, we warp the feature of the j-th view \(F^j\) into the i-th view via the two camera projection matrices \(P^i\) and \(P^j\in\mathbb{R}^{4\times4}\), producing
\begin{equation}
\label{eq:warp}
F^{ij}_{\mathrm{Warp}}
=\mathrm{Warp}\bigl(F^j,\;P^i,\;P^j,\;d^i_{\mathrm{cand}}\bigr),
\end{equation}
where Warp denotes the warping operation~\cite{xu2023unifying} and D denotes the depth dimension.
We then obtain the cost volume by computing the dot product between \(F^i\) and $F^{ij}_{\mathrm{Warp}}$:
\begin{equation}
C^i \;=\; \frac{F^i \,\otimes\, F^{ij}_{\mathrm{Warp}}}{\sqrt{C}},
\label{eq:cost_volume}
\end{equation}
where \(\otimes\) denotes element‐wise multiplication followed by summation over the channel dimension and a CNN-based upsampler.
Finally, denoting the vector of depth candidates by
\(\,G=[d_1,d_2,\dots,d_D]\in\mathbb R^D\), we obtain the depth map:
\begin{equation}
\label{eq:depth}
D^i \;=\; \operatorname{softmax}\bigl(C^i\bigr)\,G,\quad
D^i \in \mathbb{R}^{H\times W}.
\end{equation}
We follow previous research~\cite{chen2024mvsplat,charatan2024pixelsplat} to predict Gaussian parameters \(G_{\text{coarse}}\), including the Gaussian center \(\mu_i\), opacity $\alpha_i$, covariance $\Sigma_i$, and color $c_i$. Specifically, we utilize the depth map $D^i$ to unproject the 2D pixel to the 3D space location as the Gaussian center and predict opacity using a simple MLP layer. We calculate color from the predicted spherical harmonic coefficients and use a scaling matrix $s$ and a rotation matrix $R(\theta)$ to represent the covariance matrix.
\begin{equation}
\Sigma_i = R(\theta)^\top \,\mathrm{diag}(s)\,R(\theta).
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Fig_cmyk/TADC0716_cmyk.pdf}
% \vspace{-10pt}
\caption{Error maps show the intensity of inconsistency between the rendered image and the ground truth. We observe that Regions with high texture richness are often under-optimized. Therefore, we propose \textit{TADC} dynamically control the density of Gaussians according to texture richness.
}
% \vspace{-10pt}
\label{fig:tadc}
\end{figure}
\subsection{Texture-Aware Density Control}
\noindent \textbf{\textit{How to Represent Texture Richness?}}
Regions with significant color variation
(i.e., large color gradients) contain rich texture details~\cite{hu2025texture,shi2025sparse4dgs}. To provide an estimate of texture richness, we apply a high-pass filter (Sobel in this paper) to compute the first-order finite differences of pixel intensities. Given an RGB image \(I\in\mathbb{R}^{H\times W\times3}\), we compute two gradient maps by convolving \(I\) with the standard horizontal and vertical Sobel operators, respectively:
\begin{equation}
\begin{aligned}
T_x &= {I} \ast
\begin{bmatrix}
-1 & 0 & 1\\
-2 & 0 & 2\\
-1 & 0 & 1
\end{bmatrix},
T_y &= {I} \ast
\begin{bmatrix}
-1 & -2 & -1\\
0 & 0 & 0\\
1 & 2 & 1
\end{bmatrix}.
\end{aligned}
\end{equation}
We then compute the gradient magnitude, thereby deriving the texture richness map $TR = \sqrt{(T_x)^2 + (T_y)^2}$.
\noindent \textbf{\textit{Observations.}} The proposed Gaussian prediction module predicts one Gaussian primitive for each pixel. However, we observed that regions with high texture richness are often under-optimized and show large inconsistencies with the ground truth in error maps (see Fig.~\ref{fig:tadc}). More Gaussians primitives are required to fit these texture-rich areas~\cite{hu2025texture}. To this end, we propose \textit{Texture-Aware Density Control (TADC)}, a module that dynamically controls the density of Gaussians according to texture richness.
\label{sec:TADC}
\noindent \textbf{\textit{Texture Richness Perceptron for LR Images.}} To begin with, we need to recognize where the texture-rich regions are from the 2D LR images. To achieve this, we carefully designed the \textit{Texture Richness Perceptron} $\mathcal{E}_{tex}$, which is implemented using a convolutional neural network. Given an LR input image $I^{\text{LR}}$, we predict its texture richness map $\hat{TR}=\mathcal{E}_{tex}(I^{\text{LR}})$.
To optimize $\mathcal{E}_{tex}$, we employ the texture richness map $TR$ obtained as described above as supervisory signal:
\begin{equation}
\label{eq:TI_l1}
\mathcal{L}_{\mathrm{tex}}
=\mathcal{L}_1\bigl(\hat{TR},TR\bigr)\,.
\end{equation}
\noindent \textbf{\textit{Density Control.}} After obtaining the predicted Gaussian set \(G_{\text{coarse}}\) and the texture richness map $\hat{TR}$, we select the Gaussian primitives set ${G}_{den}$ exhibiting the high texture richness. Each selected Gaussian primitive is further decomposed into multiple finer primitives.
Specifically, their positions and features, denoted as \(\mathcal{X}_{\mathrm{den}}\) and \(\mathcal{F}_{\mathrm{den}}\), are input to a dedicated densification network $\mathcal{N}_{\mathrm{dens}}$ that predicts the upsampled positions and features~\cite{GenerativeDensification}. These upsampled positions and features are subsequently passed to a \emph{HEAD} module, where they are converted into fine Gaussian parameters. The set of densified Gaussian primitive is defined as:
\begin{equation}
\label{eq:dense}
G_{\mathrm{dense}} ={\textit{HEAD}}(\mathcal{N}_{\mathrm{dens}} (\mathcal{X}_{\mathrm{den}},\mathcal{F}_{\mathrm{den}}) , {G}_{den}).
\end{equation}
Finally, we obtain the refined Gaussian primitives $G_{\mathrm{refine}}$:
\begin{equation}
\label{eq:refine}
G_{\mathrm{refine}}
= G_{\mathrm{dense}} \cup G_{\mathrm{coarse}}\bigr.
\end{equation}
\subsection{Training Loss}
During training, we directly supervise the quality of the novel RGB images using Mean Squared Error (MSE) and Learned Perceptual Image Patch Similarity (LPIPS)~\cite{zhang2018unreasonable} losses. We also introduce $\mathcal{L}_{\mathrm{tex}}$ to supervise the texture richness perceptron. The total training loss is:
\begin{equation}
\label{eq:loss}
\mathcal{L} = \lambda_{\mathrm{mse}}\,\mathcal{L}_{\mathrm{mse}} + \lambda_{\mathrm{lpips}}\,\mathcal{L}_{\mathrm{lpips}} +\lambda_{\mathrm{tex}}\,\mathcal{L}_{\mathrm{tex}} ,
\end{equation}
where $\lambda_{\mathrm{mse}}$, $\lambda_{\mathrm{lpips}}$ and $\lambda_{\mathrm{tex}}$ are balancing weights.
\begin{figure*}[t]
\centering
\includegraphics[width=0.92\linewidth]{Fig_cmyk/pic_duibi1113_cmyk.pdf}
\caption{Qualitative comparison of novel views on RealEstate10k and ACID datasets. Compared with baseline methods, our approach yields sharper high-resolution novel views with cleaner fine-grained details (e.g., the chandelier and wall painting).
}
\label{fig:duibi}
\end{figure*}
\begin{table*}[t]
\centering
\scriptsize
\begin{tabular}{c|l|ccc|ccc|ccc}
\multirow{2}{*}{Datasets}& \multirow{2}{*}{Methods} &\multicolumn{3}{c|}{{1/2 \text{Resolution}}} & \multicolumn{3}{c|}{{1/4 \text{Resolution}}} & \multicolumn{3}{c}{{1/8 \text{Resolution}}} \\
\cline{3-11}
&&PSNR\(\uparrow\)&SSIM\(\uparrow\)&LPIPS\(\downarrow\)&PSNR\(\uparrow\)&SSIM\(\uparrow\)&LPIPS\(\downarrow\)&PSNR\(\uparrow\)&SSIM\(\uparrow\)&LPIPS\(\downarrow\)\\
\ChangeRT{1.2pt}
\multirow{4}{*}{RealEstate10K}&pixelSplat~\cite{charatan2024pixelsplat}& 24.56& 0.805 & 0.254 & \underline{22.80}& 0.720&0.358& \underline{20.64}& 0.618&0.491\\
&MVSplat~\cite{chen2024mvsplat}& 24.06 & 0.812& 0.177& 22.20& 0.719&0.262& 20.32& 0.614&0.375\\
&TranSplat~\cite{zhang2025transplat}& \underline{24.60} & \underline{0.827} & \underline{0.168} & 22.53& \underline{0.727}&\underline{0.256}& 20.51& \underline{0.623}& \underline{0.369} \\
&\textbf{SRSplat (Ours)}& \textbf{25.20}& \textbf{ 0.844}& \textbf{0.152}& \textbf{23.99}& \textbf{0.802}& \textbf{0.189}& \textbf{21.98}& \textbf{0.712}& \textbf{0.267}\\
\ChangeRT{1.2pt}
\multirow{4}{*}{ACID}&pixelSplat~\cite{charatan2024pixelsplat}& \underline{26.56}& 0.780 & 0.284 & \underline{24.81}& \underline{0.687}&0.405& \underline{22.97}& \underline{0.603}&0.527\\
&MVSplat~\cite{chen2024mvsplat}& 26.18 & 0.784& \underline{0.192}& 24.25& 0.675&0.282& 22.61& 0.583&\underline{0.377}\\
&TranSplat~\cite{zhang2025transplat}& 26.25 & \underline{0.786} & 0.193 & 24.34& 0.677&\underline{0.278}& 22.57& 0.582& 0.378 \\
&\textbf{{SRSplat (Ours)}}& \textbf{27.39}& \textbf{0.823}& \textbf{0.163}& \textbf{25.62}& \textbf{0.757}& \textbf{0.218}& \textbf{23.79}& \textbf{0.660}& \textbf{0.288}\\
% \cline{1-11}
% \bottomrule
\end{tabular}
\caption{Quantitative comparison under different input resolutions. SRSplat surpass all baseline methods in terms of PSNR, SSIM, and LPIPS. (\textbf{Bold} figures indicate the best and \underline{underlined} figures indicate the second best)}
\label{tab:comparison}
\end{table*}
\begin{table}[ht]
\centering
\renewcommand\arraystretch{1.2}
\scriptsize
\begin{tabular}{l|ccc|ccc}
% \toprule
\multirow{2}{*}{\textbf{{Methods}}} &\multicolumn{3}{c|}{Re10k$\rightarrow$DTU} & \multicolumn{3}{c}{Re10k$\rightarrow$ACID}\\
\cline{2-7}
&PSNR&SSIM&LPIPS&PSNR&SSIM&LPIPS \\
\ChangeRT{1.2pt}
pixelSplat & 12.66 & 0.367 & 0.577 & \underline{24.70} & \underline{0.679} & 0.416 \\
MVSplat & \underline{13.75} & \underline{0.410} & \underline{0.511} & 24.11 & 0.677 & 0.297 \\
TranSplat & 13.56 & 0.391 & 0.532 & 24.22 & 0.675 & \underline{0.296} \\
Ours& \textbf{13.80} & \textbf{0.417} & \textbf{0.445} &
\textbf{25.55} & \textbf{0.754} & \textbf{0.226} \\
\end{tabular}
\caption{{Quantative comparisons of cross-dataset generalization.} We conduct zero-shot evaluations on the ACID and DTU datasets with models trained on RealEstate10K.}
\label{tab:cross_dataset}
\end{table}
\section{Experiments}
\subsection{Experimental Settings}
\textbf{Datasets.} Our model is trained on the large-scale RealEstate10K~\cite{zhou2018stereo} and ACID~\cite{liu2021infinite} datasets. We evaluate our method on the RealEstate10K, ACID, and DTU datasets. We downsample the original training sets by factors of 2, 4, and 8 for training and evaluation. Following MVSplat~\cite{chen2024mvsplat}, the model is trained with two context views, and all methods are evaluated on three novel target views. For the DTU dataset, results are reported on 16 validation scenes, each with four novel views.
\noindent\textbf{Baselines and metrics.}
We compare SRSplat with typical methods in scene-level novel view synthesis, including pixelSplat (CVPR24)~\cite{charatan2024pixelsplat}, MVSplat (ECCV24)~\cite{chen2024mvsplat}, and TranSplat (AAAI25)~\cite{zhang2025transplat}.
We also compare the per-scene optimization method SRGS~\cite{2024srgs} and FSGS~\cite{zhu2024fsgs} on DTU dataset.
In the following sections, we report the average PSNR, SSIM~\cite{wang2004image}, and LPIPS~\cite{zhang2018unreasonable} for all baselines.
\noindent\textbf{Implementation.} We implement SRSplat using the PyTorch framework. For each scene, we downsample images by factors of 2, 4, and 8 to create LR inputs. Training is conducted with a batch size of 10 across 5 NVIDIA RTX 4090 GPUs using the Adam~\cite{kingma2014adam} optimizer for 10k iterations, requiring approximately 2 days.
% The weights of the loss function, Eq.~\ref{eq:loss}, $\lambda_{\mathrm{mse}}$, $\lambda_{\mathrm{lpips}}$, and $\lambda_{\mathrm{tex}}$ are set to 1, 0.05, and 0.05, respectively.
% More details on the implementation can be found in the supplementary materials.
\begin{figure}[h]
\centering
\includegraphics[width=0.95\linewidth]{Fig_cmyk/dtu_cmyk.pdf}
\caption{The qualitative comparisons on DTU datasets.
}
\label{fig:dtu}
\end{figure}
\subsection{Main Results}
\noindent\textbf{Quantitative Results.} The quantitative results are presented in Tab.~\ref{tab:comparison}. SRSplat achieves state-of-the-art performance across all visual quality metrics on both RealEstate10K and ACID benchmarks. Notably, as the input image resolution decreases, the reconstruction performance of previous methods degrade significantly, while SRSplat maintains comparatively strong metrics. This is because our proposed method effectively preserves and enhances high-frequency texture representations under LR inputs.
\noindent\textbf{Qualitative Results.} Fig.~\ref{fig:duibi} presents the visualization results of SRSplat and other methods. SRSplat achieves superior quality on novel view images across various challenging scenes. In the first row (large smooth regions like the open water band), our approach reconstructs sharper object boundaries (ship hull and mast) despite scarce intrinsic texture. In texture-rich regions (second row), our \textit{TADC} module adaptively densifies Gaussians, yielding cleaner texture details while other methods blur them.
\begin{table}[t]
\centering
\renewcommand\arraystretch{1.2}
\scriptsize
\begin{tabular}{l|cccc}
{\textbf{{Methods}}}
& PSNR$\uparrow$ & SSIM$\uparrow$ & LPIPS$\downarrow$ & Time$\downarrow$ \\
\ChangeRT{1.2pt}
SRGS~\cite{2024srgs} & 12.42 & 0.327 & 0.598 & \underline{300s} \\
SRGS + FSGS~\cite{zhu2024fsgs} & \underline{13.72} & \textbf{0.444} & \underline{0.481} & 420s \\
Ours & \textbf{13.80} & \underline{0.417} & \textbf{0.445} & \textbf{0.2s} \\
\end{tabular}%
\caption{{Quantitative results comparing with the per-scene optimization method.}}
\label{tab:SRGS_FSGS_com}
\end{table}
\begin{table}[t]
\centering
\renewcommand\arraystretch{1.2}
\scriptsize
\begin{tabular}{l|ccc|ccc}
% \toprule
\multirow{2}{*}{\textbf{{Methods}}} &\multicolumn{3}{c|}{Setting(a)} & \multicolumn{3}{c}{Setting(b)}\\
\cline{2-7}
&PSNR&SSIM&LPIPS&PSNR&SSIM&LPIPS\\
\ChangeRT{1.2pt}
pixelSplat & \underline{21.38} & 0.642 & 0.415 & \underline{21.63} & \underline{0.704} & 0.365 \\
MVSplat & {21.04} & {0.643} & 0.289 & 20.52 & 0.686 & 0.289 \\
TranSplat & 21.24 & \underline{0.652} & \underline{0.279} & 20.88 & 0.698 & \underline{0.286} \\
Ours& \textbf{22.07} & \textbf{0.727} & \textbf{0.219} &
\textbf{21.74} & \textbf{0.711} & \textbf{0.280} \\
\end{tabular}
\caption{{Quantative comparisons of cross-resolution generalization.} Setting (a): train on 1/8 resolution inputs with original-resolution supervision, test with 4× upsampling. Setting (b): train on 1/4 resolution inputs with 1/2 resolution supervision, test with 4× upsampling. }
\label{tab:cross_resolution}
\end{table}
\subsection{Other Results}
\noindent\textbf{Cross-dataset generalization.} To verify the cross-dataset generalization ability of SRSplat, we train the model on the RealEstate10K dataset downsampled by a factor of 4 and directly evaluate it on the ACID and DTU datasets. The quantitative results are presented in Tab.~\ref{tab:cross_dataset}. Compared with previous methods, SRSplat achieves significant improvements in generalization, obtaining +0.85 dB PSNR, +0.075 SSIM, and –0.07 LPIPS on the ACID dataset compared to suboptimal baselines. As depicted in Fig.~\ref{fig:dtu}, the images generated by SRSplat exhibit finer texture details and less blurriness.
\noindent\textbf{Cross-resolution generalization.} We assess cross-resolution generalization in two experimental configurations. First, we train on RealEstate10K images downsampled by a factor of 8, using original-resolution (1×) ground truth as supervision, and then perform 4× super-resolution during testing. Second, we train on inputs downsampled by a factor of 4, supervised by the 2×-downsampled ground truth, and then apply 4× super-resolution during evaluation. As shown in Tab.~\ref{tab:cross_resolution}, SRSplat exhibits superior cross-resolution generalization compared to baselines.
\noindent\textbf{Comparison with the per-scene optimization method.} We further compare SRSplat with the representative super-resolution method SRGS~\cite{2024srgs} and an enhanced version of SRGS with FSGS~\cite{zhu2024fsgs} on the DTU dataset. Qualitative results are shown in Fig.~\ref{fig:dtu} and the quantitative results are reported in Tab.~\ref{tab:cross_dataset}. Compared to per-scene optimization methods, SRSplat not only achieves competitive performance but also enables real-time reconstruction.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Fig_cmyk/TR_ablation0726_cmyk.pdf}
\caption{Comparison of texture‐richness (TR) maps. (a) TR map obtained by applying the Sobel to the HR GT image; (b) TR map extracted by Sobel on the LR image; (c) TR map perceived by our TR perceptron from the LR image.
}
\label{fig:TR_ablation}
\end{figure}
\subsection{Ablation Study}
We conduct a detailed ablation study of SRSplat on the RealEstate10K datasets, downsampled by a factor of 4.
\noindent\textbf{Importance of each module of SRSplat.} In this section, we analyze the effectiveness of the proposed modules in detail. The \textit{RGFE} module compensates for the lack of high-frequency information in LR input images by leveraging external reference images. The \textit{TADC} adaptively densifies Gaussian primitives based on the texture richness of internal images. The experimental results are summarized in Tab.~\ref{tab:Ablation} (rows 1–3), demonstrating the pivotal roles of these modules in super-resolution reconstruction.
\noindent\textbf{Importance of texture richness perceptron.} The texture richness (TR) perceptron plays a crucial role in distilling texture richness from LR inputs, which is essential for subsequent texture-aware density control. When we remove TR perceptron and use the Sobel operator directly to obtain TR maps from LR images, we observe a decrease in reconstruction quality, as shown in Tab.~\ref{tab:Ablation} (row 4). Fig.~\ref{fig:TR_ablation} presents a comparison of TR maps extracted by different methods. The maps produced by the TR perceptron align more accurately with scene texture details (e.g., the curtain rod).
% \noindent\textbf{Ablation of multi-view number.} To assess SRSplat's multi-view capability, we train the model on RealEstate10K with 2-view inputs and directly evaluate it on the DTU dataset with more view inputs. Compared with the 2-view results presented in Tab.~\ref{tab:cross_dataset}, SRSplat achieves a PSNR of 13.91, an SSIM of 0.435, and an LPIPS of 0.446 with 3-view inputs. This indicates that SRSplat benefits from additional input views when they are available at test time.
% \noindent\textbf{Sensitivity Analysis of Hyper-parameters. }The results can be found in the supplementary materials.
\begin{table}[t]
\centering
\resizebox{0.4\textwidth}{!}{%
\renewcommand\arraystretch{1.2}
\scriptsize
\begin{tabular}{cc|ccc}
\textit{FGFE} & \textit{TADC}
& PSNR$\uparrow$ & SSIM$\uparrow$ & LPIPS$\downarrow$ \\
\ChangeRT{1.2pt}
& & 22.20 & 0.719 & 0.262 \\
\checkmark & & {23.56} & {0.782} & {0.207} \\
\checkmark & \checkmark & \textbf{23.99} & \textbf{0.802} & \textbf{0.189} \\
\ChangeRT{1.2pt}
\multicolumn{2}{c|}{w/o TR perceptron}
&23.75 & 0.795 & 0.192
\end{tabular}%
}
\caption{Ablation study on RealEstate10K.}
\label{tab:Ablation}
\end{table}
\section{Conclusion}
In this work, we propose \textbf{SRSplat}, the first feed-forward Super-Resolution framework. Our method initially constructs a reference gallery tailored to the unique characteristics of each scene. With this reference gallery, we introduce \textit{RGFE} to align and fuse features from the LR input images and their reference. To further refine scene details, we propose \textit{TADC}, which adaptively adjusts the density of scene’s Gaussians according to the texture richness. Extensive experiments on multiple datasets demonstrate that SRSplat outperforms existing methods.
\section*{Acknowledgments}
This work was supported in part by the National Natural Science Foundation of China under Grants (No. 62206082, 422204, 62502135), the Key
Research and Development Program of Zhejiang Province
(No. 2025C01026), the Zhejiang Provincial Natural Science Foundation of China under Grants
(No. LQN25F030014). This research was also supported by the National College Student Innovation and Entrepreneurship Training Program of China under Grants (No. 202410336018 and 202510336016).
\bibliography{main}
\end{document}
\begin{thebibliography}{43}
\providecommand{\natexlab}[1]{#1}
\bibitem[{Achiam et~al.(2023)Achiam, Adler, Agarwal, Ahmad, Akkaya, Aleman, Almeida, Altenschmidt, Altman, Anadkat et~al.}]{achiam2023gpt}
Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F.~L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et~al. 2023.
\newblock Gpt-4 technical report.
\newblock \emph{arXiv preprint arXiv:2303.08774}.
\bibitem[{B{\"o}siger et~al.(2024)B{\"o}siger, Dusmanu, Pollefeys, and Bauer}]{bosiger2024mariner}
B{\"o}siger, L.; Dusmanu, M.; Pollefeys, M.; and Bauer, Z. 2024.
\newblock MaRINeR: Enhancing Novel Views by Matching Rendered Images with Nearby References.
\newblock In \emph{European Conference on Computer Vision}, 76--94. Springer.
\bibitem[{Cao, Ren, and Fu(2022)}]{cao2022mvsformer}
Cao, C.; Ren, X.; and Fu, Y. 2022.
\newblock MVSFormer: Multi-view stereo by learning robust image features and temperature-based depth.
\newblock \emph{arXiv preprint arXiv:2208.02541}.
\bibitem[{Cao et~al.(2022)Cao, Liang, Zhang, Li, Zhang, Wang, and Gool}]{cao2022reference}
Cao, J.; Liang, J.; Zhang, K.; Li, Y.; Zhang, Y.; Wang, W.; and Gool, L.~V. 2022.
\newblock Reference-based image super-resolution with deformable attention transformer.
\newblock In \emph{European conference on computer vision}, 325--342. Springer.
\bibitem[{Charatan et~al.(2024)Charatan, Li, Tagliasacchi, and Sitzmann}]{charatan2024pixelsplat}
Charatan, D.; Li, S.~L.; Tagliasacchi, A.; and Sitzmann, V. 2024.
\newblock pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction.
\newblock In \emph{Proceedings of the IEEE/CVF conference on computer vision and pattern recognition}, 19457--19467.
\bibitem[{Chen et~al.(2022)Chen, Xu, Geiger, Yu, and Su}]{chen2022tensorf}
Chen, A.; Xu, Z.; Geiger, A.; Yu, J.; and Su, H. 2022.
\newblock Tensorf: Tensorial radiance fields.
\newblock In \emph{European conference on computer vision}, 333--350. Springer.
\bibitem[{Chen et~al.(2024{\natexlab{a}})Chen, Xu, Zheng, Zhuang, Pollefeys, Geiger, Cham, and Cai}]{chen2024mvsplat}
Chen, Y.; Xu, H.; Zheng, C.; Zhuang, B.; Pollefeys, M.; Geiger, A.; Cham, T.-J.; and Cai, J. 2024{\natexlab{a}}.
\newblock Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images.
\newblock In \emph{European Conference on Computer Vision}, 370--386. Springer.
\bibitem[{Chen et~al.(2024{\natexlab{b}})Chen, Zheng, Xu, Zhuang, Vedaldi, Cham, and Cai}]{chen2024mvsplat360}
Chen, Y.; Zheng, C.; Xu, H.; Zhuang, B.; Vedaldi, A.; Cham, T.-J.; and Cai, J. 2024{\natexlab{b}}.
\newblock Mvsplat360: Feed-forward 360 scene synthesis from sparse views.
\newblock \emph{Advances in Neural Information Processing Systems}, 37: 107064--107086.
\bibitem[{Fei et~al.(2024)Fei, Zheng, Duan, Zhan, Tomizuka, Keutzer, and Lu}]{fei2024pixelgaussian}
Fei, X.; Zheng, W.; Duan, Y.; Zhan, W.; Tomizuka, M.; Keutzer, K.; and Lu, J. 2024.
\newblock Pixelgaussian: Generalizable 3d gaussian reconstruction from arbitrary views.
\newblock \emph{arXiv preprint arXiv:2410.18979}.
\bibitem[{Feng et~al.(2024)Feng, He, Wang, Yang, Li, Chen, Kuang, Fan, Jun et~al.}]{2024srgs}
Feng, X.; He, Y.; Wang, Y.; Yang, Y.; Li, W.; Chen, Y.; Kuang, Z.; Fan, J.; Jun, Y.; et~al. 2024.
\newblock SRGS: Super-Resolution 3D Gaussian Splatting.
\newblock \emph{arXiv preprint arXiv:2404.10318}.
\bibitem[{Hu et~al.(2025)Hu, Shi, Yang, Chen, Gu, Ding, He, and Fan}]{hu2025texture}
Hu, X.; Shi, C.; Yang, C.; Chen, M.; Gu, X.; Ding, J.; He, J.; and Fan, J. 2025.
\newblock Texture-aware 3d Gaussian splatting for sparse view reconstructions.
\newblock \emph{Applied Soft Computing}, 113530.
\bibitem[{Huang et~al.(2023{\natexlab{a}})Huang, Mees, Zeng, and Burgard}]{huang2023visual}
Huang, C.; Mees, O.; Zeng, A.; and Burgard, W. 2023{\natexlab{a}}.
\newblock Visual language maps for robot navigation.
\newblock In \emph{2023 IEEE International Conference on Robotics and Automation (ICRA)}, 10608--10615. IEEE.
\bibitem[{Huang et~al.(2023{\natexlab{b}})Huang, Li, Hu, Chen, and Wang}]{huang2023refsr}
Huang, X.; Li, W.; Hu, J.; Chen, H.; and Wang, Y. 2023{\natexlab{b}}.
\newblock Refsr-nerf: Towards high fidelity and super resolution view synthesis.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 8244--8253.
\bibitem[{Jiang et~al.(2021)Jiang, Chan, Wang, Loy, and Liu}]{jiang2021robust}
Jiang, Y.; Chan, K.~C.; Wang, X.; Loy, C.~C.; and Liu, Z. 2021.
\newblock Robust reference-based super-resolution via c2-matching.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 2103--2112.
\bibitem[{Kerbl et~al.(2023)Kerbl, Kopanas, Leimk{\"u}hler, and Drettakis}]{kerbl20233d}
Kerbl, B.; Kopanas, G.; Leimk{\"u}hler, T.; and Drettakis, G. 2023.
\newblock 3D Gaussian Splatting for Real-Time Radiance Field Rendering.
\newblock \emph{ACM Transactions on Graphics}, 42(4).
\bibitem[{Kingma(2014)}]{kingma2014adam}
Kingma, D.~P. 2014.
\newblock Adam: A method for stochastic optimization.
\newblock \emph{arXiv preprint arXiv:1412.6980}.
\bibitem[{Labs et~al.(2025)Labs, Batifol, Blattmann, Boesel, Consul, Diagne, Dockhorn, English, English, Esser, Kulal, Lacey, Levi, Li, Lorenz, Müller, Podell, Rombach, Saini, Sauer, and Smith}]{labs2025flux1kontextflowmatching}
Labs, B.~F.; Batifol, S.; Blattmann, A.; Boesel, F.; Consul, S.; Diagne, C.; Dockhorn, T.; English, J.; English, Z.; Esser, P.; Kulal, S.; Lacey, K.; Levi, Y.; Li, C.; Lorenz, D.; Müller, J.; Podell, D.; Rombach, R.; Saini, H.; Sauer, A.; and Smith, L. 2025.
\newblock FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space.
\newblock arXiv:2506.15742.
\bibitem[{Liang et~al.(2021)Liang, Cao, Sun, Zhang, Van~Gool, and Timofte}]{liang2021swinir}
Liang, J.; Cao, J.; Sun, G.; Zhang, K.; Van~Gool, L.; and Timofte, R. 2021.
\newblock Swinir: Image restoration using swin transformer.
\newblock In \emph{Proceedings of the IEEE/CVF international conference on computer vision}, 1833--1844.
\bibitem[{Liu et~al.(2021{\natexlab{a}})Liu, Tucker, Jampani, Makadia, Snavely, and Kanazawa}]{liu2021infinite}
Liu, A.; Tucker, R.; Jampani, V.; Makadia, A.; Snavely, N.; and Kanazawa, A. 2021{\natexlab{a}}.
\newblock Infinite nature: Perpetual view generation of natural scenes from a single image.
\newblock In \emph{Proceedings of the IEEE/CVF International Conference on Computer Vision}, 14458--14467.
\bibitem[{Liu et~al.(2021{\natexlab{b}})Liu, Lin, Cao, Hu, Wei, Zhang, Lin, and Guo}]{liu2021swin}
Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021{\natexlab{b}}.
\newblock Swin transformer: Hierarchical vision transformer using shifted windows.
\newblock In \emph{Proceedings of the IEEE/CVF international conference on computer vision}, 10012--10022.
\bibitem[{Lu et~al.(2021)Lu, Li, Tao, Lu, and Jia}]{lu2021masa}
Lu, L.; Li, W.; Tao, X.; Lu, J.; and Jia, J. 2021.
\newblock Masa-sr: Matching acceleration and spatial adaptation for reference-based image super-resolution.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 6368--6377.
\bibitem[{Mildenhall et~al.(2021)Mildenhall, Srinivasan, Tancik, Barron, Ramamoorthi, and Ng}]{mildenhall2021nerf}
Mildenhall, B.; Srinivasan, P.~P.; Tancik, M.; Barron, J.~T.; Ramamoorthi, R.; and Ng, R. 2021.
\newblock Nerf: Representing scenes as neural radiance fields for view synthesis.
\newblock \emph{Communications of the ACM}, 65(1): 99--106.
\bibitem[{M{\"u}ller et~al.(2022)M{\"u}ller, Evans, Schied, and Keller}]{muller2022instant}
M{\"u}ller, T.; Evans, A.; Schied, C.; and Keller, A. 2022.
\newblock Instant neural graphics primitives with a multiresolution hash encoding.
\newblock \emph{ACM transactions on graphics (TOG)}, 41(4): 1--15.
\bibitem[{Nam et~al.(2024)Nam, Sun, Kang, Lee, Oh, and Park}]{GenerativeDensification}
Nam, S.; Sun, X.; Kang, G.; Lee, Y.; Oh, S.; and Park, E. 2024.
\newblock Generative Densification: Learning to Densify Gaussians for High-Fidelity Generalizable 3D Reconstruction.
\newblock \emph{arXiv preprint arXiv:2412.06234}.
\bibitem[{Shi et~al.(2025{\natexlab{a}})Shi, Chen, Mao, Yang, Hu, Ding, and Yu}]{shi2025realm}
Shi, C.; Chen, M.; Mao, Y.; Yang, C.; Hu, X.; Ding, J.; and Yu, Z. 2025{\natexlab{a}}.
\newblock REALM: An MLLM-Agent Framework for Open World 3D Reasoning Segmentation and Editing on Gaussian Splatting.
\newblock \emph{arXiv preprint arXiv:2510.16410}.
\bibitem[{Shi et~al.(2025{\natexlab{b}})Shi, Yang, Hu, Chen, Pan, Yang, Ding, Yu, and Yu}]{shi2025sparse4dgs}
Shi, C.; Yang, C.; Hu, X.; Chen, M.; Pan, W.; Yang, Y.; Ding, J.; Yu, Z.; and Yu, J. 2025{\natexlab{b}}.
\newblock Sparse4DGS: 4D Gaussian Splatting for Sparse-Frame Dynamic Scene Reconstruction.
\newblock \emph{arXiv preprint arXiv:2511.07122}.
\bibitem[{Shi et~al.(2025{\natexlab{c}})Shi, Yang, Hu, Yang, Ding, and Tan}]{shi2025mmgs}
Shi, C.; Yang, C.; Hu, X.; Yang, Y.; Ding, J.; and Tan, M. 2025{\natexlab{c}}.
\newblock MMGS: Multi-Model Synergistic Gaussian Splatting for Sparse View Synthesis.
\newblock \emph{Image and Vision Computing}, 158: 105512.
\bibitem[{Sun et~al.(2024)Sun, Li, Liu, Chen, Pei, Zou, Yan, and Yang}]{sun2024coser}
Sun, H.; Li, W.; Liu, J.; Chen, H.; Pei, R.; Zou, X.; Yan, Y.; and Yang, Y. 2024.
\newblock Coser: Bridging image and language for cognitive super-resolution.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 25868--25878.
\bibitem[{Tang et~al.(2024)Tang, Ye, Ye, Lin, Zhou, Chen, and Ouyang}]{tang2024hisplat}
Tang, S.; Ye, W.; Ye, P.; Lin, W.; Zhou, Y.; Chen, T.; and Ouyang, W. 2024.
\newblock Hisplat: Hierarchical 3d gaussian splatting for generalizable sparse-view reconstruction.
\newblock \emph{arXiv preprint arXiv:2410.06245}.
\bibitem[{Tian et~al.(2025)Tian, Tan, Xie, and Ma}]{tian2025drivingforward}
Tian, Q.; Tan, X.; Xie, Y.; and Ma, L. 2025.
\newblock Drivingforward: Feed-forward 3d gaussian splatting for driving scene reconstruction from flexible surround-view input.
\newblock In \emph{Proceedings of the AAAI Conference on Artificial Intelligence}, volume~39, 7374--7382.
\bibitem[{Wan et~al.(2025)Wan, Cheng, Shao, and Zuo}]{wan2025s2gaussian}
Wan, Y.; Cheng, Y.; Shao, M.; and Zuo, W. 2025.
\newblock S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splatting.
\newblock \emph{CVPR}.
\bibitem[{Wang et~al.(2021)Wang, Wu, Guo, Zhang, Tai, and Hu}]{wang2021nerf-sr}
Wang, C.; Wu, X.; Guo, Y.-C.; Zhang, S.-H.; Tai, Y.-W.; and Hu, S.-M. 2021.
\newblock NeRF-SR: High-Quality Neural Radiance Fields using Super-Sampling.
\newblock \emph{arXiv}.
\bibitem[{Wang et~al.(2004)Wang, Bovik, Sheikh, and Simoncelli}]{wang2004image}
Wang, Z.; Bovik, A.~C.; Sheikh, H.~R.; and Simoncelli, E.~P. 2004.
\newblock Image quality assessment: from error visibility to structural similarity.
\newblock \emph{IEEE transactions on image processing}, 13(4): 600--612.
\bibitem[{Xu et~al.(2025)Xu, Peng, Wang, Blum, Barath, Geiger, and Pollefeys}]{xu2025depthsplat}
Xu, H.; Peng, S.; Wang, F.; Blum, H.; Barath, D.; Geiger, A.; and Pollefeys, M. 2025.
\newblock Depthsplat: Connecting gaussian splatting and depth.
\newblock In \emph{Proceedings of the Computer Vision and Pattern Recognition Conference}, 16453--16463.
\bibitem[{Xu et~al.(2022)Xu, Zhang, Cai, Rezatofighi, and Tao}]{xu2022gmflow}
Xu, H.; Zhang, J.; Cai, J.; Rezatofighi, H.; and Tao, D. 2022.
\newblock Gmflow: Learning optical flow via global matching.
\newblock In \emph{Proceedings of the IEEE/CVF conference on computer vision and pattern recognition}, 8121--8130.
\bibitem[{Xu et~al.(2023)Xu, Zhang, Cai, Rezatofighi, Yu, Tao, and Geiger}]{xu2023unifying}
Xu, H.; Zhang, J.; Cai, J.; Rezatofighi, H.; Yu, F.; Tao, D.; and Geiger, A. 2023.
\newblock Unifying flow, stereo and depth estimation.
\newblock \emph{IEEE Transactions on Pattern Analysis and Machine Intelligence}, 45(11): 13941--13958.
\bibitem[{Yang et~al.(2020)Yang, Yang, Fu, Lu, and Guo}]{yang2020learning}
Yang, F.; Yang, H.; Fu, J.; Lu, H.; and Guo, B. 2020.
\newblock Learning texture transformer network for image super-resolution.
\newblock In \emph{Proceedings of the IEEE/CVF conference on computer vision and pattern recognition}, 5791--5800.
\bibitem[{Yang et~al.(2024)Yang, Kang, Huang, Xu, Feng, and Zhao}]{yang2024depth}
Yang, L.; Kang, B.; Huang, Z.; Xu, X.; Feng, J.; and Zhao, H. 2024.
\newblock Depth anything: Unleashing the power of large-scale unlabeled data.
\newblock In \emph{Proceedings of the IEEE/CVF conference on computer vision and pattern recognition}, 10371--10381.
\bibitem[{Yao et~al.(2018)Yao, Luo, Li, Fang, and Quan}]{yao2018mvsnet}
Yao, Y.; Luo, Z.; Li, S.; Fang, T.; and Quan, L. 2018.
\newblock Mvsnet: Depth inference for unstructured multi-view stereo.
\newblock In \emph{Proceedings of the European conference on computer vision (ECCV)}, 767--783.
\bibitem[{Zhang et~al.(2025)Zhang, Zou, Li, Yi, and Wang}]{zhang2025transplat}
Zhang, C.; Zou, Y.; Li, Z.; Yi, M.; and Wang, H. 2025.
\newblock Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers.
\newblock In \emph{Proceedings of the AAAI Conference on Artificial Intelligence}, volume~39, 9869--9877.
\bibitem[{Zhang et~al.(2018)Zhang, Isola, Efros, Shechtman, and Wang}]{zhang2018unreasonable}
Zhang, R.; Isola, P.; Efros, A.~A.; Shechtman, E.; and Wang, O. 2018.
\newblock The unreasonable effectiveness of deep features as a perceptual metric.
\newblock In \emph{Proceedings of the IEEE conference on computer vision and pattern recognition}, 586--595.
\bibitem[{Zhou et~al.(2018)Zhou, Tucker, Flynn, Fyffe, and Snavely}]{zhou2018stereo}
Zhou, T.; Tucker, R.; Flynn, J.; Fyffe, G.; and Snavely, N. 2018.
\newblock Stereo magnification: Learning view synthesis using multiplane images.
\newblock \emph{arXiv preprint arXiv:1805.09817}.
\bibitem[{Zhu et~al.(2024)Zhu, Fan, Jiang, and Wang}]{zhu2024fsgs}
Zhu, Z.; Fan, Z.; Jiang, Y.; and Wang, Z. 2024.
\newblock Fsgs: Real-time few-shot view synthesis using gaussian splatting.
\newblock In \emph{European conference on computer vision}, 145--163. Springer.
\end{thebibliography}