Landmark-4K: UHD RefSR Dataset Benchmark

• Landmark-4K is a reference-based image super-resolution dataset designed for UHD landmark imagery, offering an average resolution of 3295×3295 pixels and controlled viewpoint consistency.
• It curates high-quality images into self-reference and cross-reference pairs to benchmark restoration of fine textures under real-world degradation conditions.
• Benchmark evaluations on Landmark-4K distinctly measure performance across PSNR, SSIM, LPIPS, and DISTS, validating its role in advancing RefSR methodologies.

Landmark-4K is the first Reference-based Image Super-Resolution (RefSR) dataset specifically designed for Ultra-High-Definition (UHD) landmark scenarios. It was introduced to address two stated deficiencies of prior RefSR datasets: relatively small image sizes and insufficient fine-grained detail in both target and reference images. The dataset contains 185 high-quality images covering 49 landmark categories, with an average resolution of $3295\times3295$ pixels, and is structured so that low-resolution target images can be paired with high-resolution reference images for both self-reference and cross-reference evaluation under real-world degradation (Shi et al., 14 Aug 2025).

1. Definition and research motivation

In RefSR, a low-resolution image is restored by using semantic and texture information from an additional reference high-resolution image. Landmark-4K was created for this setting, with an explicit emphasis on UHD landmark imagery rather than the lower-resolution image regimes that dominated earlier RefSR benchmarks. The motivating claim in the originating paper is that modern devices routinely capture images at UHD resolutions, making high-resolution, realistic benchmarks more relevant and necessary for contemporary super-resolution research (Shi et al., 14 Aug 2025).

The dataset is therefore not a generic landmark corpus. Its design is task-specific: it is intended for benchmarking methods that must recover fine structures and textures by transferring information from a related but not necessarily identical reference view. This framing is important because the dataset targets a harder operating point than conventional single-image SR benchmarks. The reference image is useful only insofar as a method can align viewpoint-consistent content and selectively fuse recoverable detail.

A common misconception is to treat Landmark-4K as simply a larger variant of existing RefSR datasets. The evidence provided in the paper suggests a more specific distinction: the dataset is organized around UHD scale, real-world degradation, and controlled inter-image viewpoint consistency, all of which directly affect whether reference information is usable in practice (Shi et al., 14 Aug 2025).

2. Data acquisition and curation procedure

The construction pipeline begins with an initial collection of approximately 1,000 landmark images gathered from public online sources, specifically pexels.com. Resolution pre-filtering then excludes any image below 2K resolution. Images above 4096 pixels on either side are downscaled proportionally so that the maximum side is 4096 pixels, which the paper describes as a balance between detail and computational practicality (Shi et al., 14 Aug 2025).

Quality curation removes images with blur or noticeable noise. The images are then grouped by landmark type, and images with strongly differing viewpoints within a category are excluded to ensure reference/target consistency. After this curation process, the final dataset contains 185 high-quality images covering 49 landmark categories, with each category containing 3–4 distinct images from various controlled viewpoints (Shi et al., 14 Aug 2025).

This organization is central to the dataset’s experimental semantics. For each category, one image acts as the self-reference, while the remaining images enable cross-reference generalization tests. In other words, Landmark-4K does not merely supply isolated HR images; it supplies small viewpoint-controlled sets that instantiate the reference relation required by RefSR. This suggests that the curation criterion is as important as raw resolution, because reference usefulness depends on viewpoint compatibility as much as on texture richness.

3. Dataset composition, UHD scale, and pairing structure

Landmark-4K focuses on famous landmarks from around the globe, captured under various conditions, with an emphasis on semantic and textural richness. The dataset’s average image resolution is $3295\times3295$, and its pairing protocol creates target low-resolution images through degradation of HR images while retaining reference HR images captured from slightly different, but consistent, viewpoints (Shi et al., 14 Aug 2025).

The following comparison, reported in the paper, situates Landmark-4K against earlier RefSR datasets:

Dataset	Images	Average resolution
CUFED5	126	$418\times418$
WR-SR	80	$770\times770$
Landmark-4K	185	$3295\times3295$

The dataset supports both self-reference and cross-reference evaluation. Self-reference uses the same view as reference, while cross-reference uses a different image from the same landmark category. This distinction matters because self-reference largely probes restoration capacity under ideal alignment, whereas cross-reference probes generalization under viewpoint variation.

The paper also emphasizes that Landmark-4K is large enough to support tile-based inference and patch-wise attention mechanisms pivotal in current high-resolution SR models. Its rich, distinctive textures and structures are intended to test whether a method can perform true reference-based detail transfer rather than merely hallucinating plausible high-frequency content. In that sense, Landmark-4K is designed not only as a benchmark for fidelity, but also as a stress test for alignment and fusion.

4. Real-world degradation model and evaluation protocol

To simulate real-world conditions, Landmark-4K applies degradation using the Real-ESRGAN pipeline with $\times 4$ down-sampling, rather than simple bicubic downsampling (Wang et al., 2021). The motivation given in the paper is that this mirrors complex and unknown degradations encountered in practice. Consequently, the low-resolution targets are not clean synthetic resizes; they embed the more difficult artifact structure typical of real-world restoration settings (Shi et al., 14 Aug 2025).

Evaluation on Landmark-4K uses PSNR, SSIM, LPIPS, and DISTS. The paper explicitly notes that FID is not reported for Landmark-4K because its calculation requires downsampling, which is not appropriate for UHD images (Shi et al., 14 Aug 2025). This point addresses a recurrent misunderstanding in image-generation and restoration evaluation: omission of FID here is methodological rather than incidental.

The benchmark results reported in the paper distinguish both classical and recent RefSR systems. StableSR attains PSNR 24.90, SSIM 0.722, LPIPS 0.308, and DISTS 0.153; SUPIR attains 24.44, 0.687, 0.312, and 0.137; C2-Matching attains 25.63, 0.721, 0.578, and 0.272; DATSR attains 25.64, 0.722, 0.597, and 0.269; ReFIR attains 25.21, 0.735, 0.280, and 0.149; and TriFlowSR attains 25.89, 0.776, 0.230, and 0.115, which are the best reported results across all four metrics on this benchmark (Shi et al., 14 Aug 2025).

These results indicate that Landmark-4K is discriminative enough to separate methods not only in distortion-oriented metrics such as PSNR and SSIM, but also in perceptual metrics such as LPIPS and DISTS. The paper also reports that Landmark-4K supports ablation studies on reference alignment and fusion strategies, with measurable metric gains under different combinations of the reference branch, matching, and warping. A plausible implication is that the dataset is particularly useful for studying the failure modes of misalignment-aware UHD RefSR systems.

5. Methodological context: TriFlowSR and explicit reference matching

Landmark-4K was introduced together with TriFlowSR, a framework described as explicitly achieving pattern matching between the low-resolution image and the reference high-resolution image. Within that framework, the paper designs a Reference Matching Strategy to effectively match the LR image with the reference HR image in UHD scenarios under real-world degradation (Shi et al., 14 Aug 2025).

A core component benchmarked on Landmark-4K is Patch-Ref Attention, defined as

$$\text{PatchRefAttn} = \text{softmax}\left(\frac{Q^{SR}\left[K^{SR}, K^{LR}, K^{Ref}\right]^T}{\sqrt{d_k}}\right)\cdot \left[V^{SR}, V^{LR}, V^{Ref}\right].$$

The training loss for reference fusion is written as

$$\mathcal{L}_{\psi} = \int_{0}^{1} \mathbb{E}\left[ \left\| v_{\psi}(z^{HR}_t, z^{LR}, z^{Ref}, t) - (\epsilon - z^{HR}) \right\|^2 \right] dt,$$

with

$z^{HR}_t = (1-t)z^{HR} + t\epsilon.$

The paper states that these equations are utilized in Landmark-4K experiments to provide rigorous, quantitative evaluation of SR methods’ ability to integrate reference detail in a UHD context (Shi et al., 14 Aug 2025).

Although these formulas belong to the originating method rather than to the dataset per se, they clarify the kind of algorithmic behavior Landmark-4K is meant to expose. The benchmark is not limited to testing generic upscaling capacity; it specifically interrogates explicit pattern matching, cross-branch attention, and reference-guided restoration under large images and realistic degradation. This suggests that Landmark-4K is most informative when used with methods whose internal design makes reference alignment an explicit computational object.

6. Position within the broader UHD benchmark landscape

Landmark-4K occupies a specific niche within the expanding set of 4K and UHD computer-vision benchmarks. It differs from 4K-HAZE, which provides over 30,000 pairs of 4K-resolution hazy and haze-free images for dehazing (Zheng et al., 2023), and from 4KLSDB, which contains 129,484 native 4K images for image restoration and generation, together with distinct validation and test sets of 2,000 and 1,984 images (Zhu et al., 23 May 2026). It also differs from TerraSky3D, which comprises approximately 50,000 images divided into 150 ground, aerial, and mixed scenes for multi-view 3D reconstruction of European landmarks (D'Urso et al., 30 Mar 2026).

These comparisons clarify that Landmark-4K is not primarily a generic 4K image corpus, a dehazing dataset, or a geometric reconstruction benchmark. It is a task-specific RefSR benchmark whose central innovation lies in combining UHD image scale, landmark-centric semantic richness, and reference-consistent viewpoint variation. A plausible implication is that its smaller absolute size relative to very large 4K datasets is offset by tighter alignment between dataset design and the reference-based restoration problem.

The paper’s conclusion is therefore best understood in methodological terms. Landmark-4K enables rigorous quantitative and qualitative assessment of modern RefSR methods in a setting where high-resolution detail, realistic degradation, and viewpoint-consistent references are all simultaneously present (Shi et al., 14 Aug 2025). For research on UHD RefSR, that combination is the dataset’s defining contribution.