\subsection{Qualitative Analysis} \label{sec:qualitative} Figure~\ref{fig:qualitative} presents a qualitative example of the proposed $E_v$-guided restoration. Although the adversarial image remains visually similar to its clean counterpart, the restored image exhibits noticeable high-frequency artifacts and color variations. Nevertheless, the main semantic structures, including the object's shape and fine-grained visual characteristics, remain recognizable. This observation highlights an important distinction between perceptual fidelity and adversarial robustness. Unlike conventional image reconstruction methods, our approach does not explicitly minimize the pixel-level distance to the clean image. Instead, it modifies the input according to the local even-response energy of the frozen visual encoder. Consequently, improved downstream robustness does not necessarily imply improved visual fidelity. Together with the quantitative results on CLIP and LLaVA, this example suggests that visually distorted inputs can still provide useful task-relevant information for vision-language models. \begin{figure*}[t] \centering \begin{minipage}{0.25\textwidth} \centering \includegraphics[width=\linewidth]{figures/clean.png} \small (a) Clean \end{minipage} \hfill \begin{minipage}{0.25\textwidth} \centering \includegraphics[width=\linewidth]{figures/adv.png} \small (b) Adversarial \end{minipage} \hfill \begin{minipage}{0.25\textwidth} \centering \includegraphics[width=\linewidth]{figures/restored.png} \small (c) Restored \end{minipage} \caption{ Qualitative visualization of $E_v$-guided restoration. The restored image preserves recognizable semantic structures while introducing visible high-frequency artifacts. Our defense aims to improve task-level robustness rather than pixel-level reconstruction fidelity. } \label{fig:qualitative} \end{figure*}

\subsection{Qualitative Analysis} \label{sec:qualitative} Figure~\ref{fig:qualitative} presents a qualitative example of the proposed $E_v$-guided restoration. Although the adversarial image remains visually similar to its clean counterpart, the restored image exhibits noticeable high-frequency artifacts and color variations. Nevertheless, the main semantic structures, including the object's shape and fine-grained visual characteristics, remain recognizable. This observation highlights an important distinction between perceptual fidelity and adversarial robustness. Unlike conventional image reconstruction methods, our approach does not explicitly minimize the pixel-level distance to the clean image. Instead, it modifies the input according to the local even-response energy of the frozen visual encoder. Consequently, improved downstream robustness does not necessarily imply improved visual fidelity. Together with the quantitative results on CLIP and LLaVA, this example suggests that visually distorted inputs can still provide useful task-relevant information for vision-language models. \begin{figure*}[t] \centering \begin{minipage}{0.25\textwidth} \centering \includegraphics[width=\linewidth]{figures/clean.png} \small (a) Clean \end{minipage} \hfill \begin{minipage}{0.25\textwidth} \centering \includegraphics[width=\linewidth]{figures/adv.png} \small (b) Adversarial \end{minipage} \hfill \begin{minipage}{0.25\textwidth} \centering \includegraphics[width=\linewidth]{figures/restored.png} \small (c) Restored \end{minipage} \caption{ Qualitative visualization of $E_v$-guided restoration. The restored image preserves recognizable semantic structures while introducing visible high-frequency artifacts. Our defense aims to improve task-level robustness rather than pixel-level reconstruction fidelity. } \label{fig:qualitative} \end{figure*}