RASR-Flickr30 Benchmark

• RASR-Flickr30 is a public benchmark that redefines reference-based super resolution by using per-category databases for open-world retrieval.
• It organizes 30 animal species into structured HQ/LQ pairs and reference sets to simulate real-world scenarios such as mobile zoo or museum photography.
• The baseline system, RASRNet, combines latent diffusion restoration with semantic retrieval, achieving measurable gains in metrics like PSNR and LPIPS.

RASR-Flickr30 is the first public benchmark specifically designed for Retrieval-Augmented Super Resolution (RASR). Introduced together with RASRNet, it departs from prior Reference-based Super Resolution settings that rely on manually curated target-reference image pairs by providing per-category reference databases that support open-world retrieval. The benchmark focuses on 30 animal species, with each species serving both as a target class for low-quality/high-quality (LQ/HQ) pairs and as a category for a standalone reference database (Yan et al., 13 Aug 2025).

1. Benchmark definition and design rationale

RASR-Flickr30 was constructed to study a practical RefSR regime in which a system is given only a low-quality input and must retrieve semantically relevant high-resolution references from a database before restoration. In the formulation introduced with the benchmark, the task is to restore $x_{LQ}$ conditioned on retrieved references rather than on a manually paired exemplar:

$$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$

where $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$ is the retrieval function (Yan et al., 13 Aug 2025).

The dataset design reflects this problem definition. Unlike prior datasets with fixed target-reference pairs, RASR-Flickr30 provides per-category reference databases to support open-world retrieval. The stated motivation is practical deployment in settings such as mobile photos taken in environments like zoos or museums, where category-specific reference data can be readily collected or pre-curated. This framing places the benchmark at the intersection of semantic retrieval, reference-based image restoration, and diffusion-based super-resolution, with the retrieval stage treated as a first-class component rather than a preprocessing convenience.

2. Dataset construction, composition, and preprocessing

The source of the imagery is Flickr.com, with collection prioritizing high-resolution, aesthetically pleasing animal photographs. Filtering criteria were visual clarity, photographic quality, and correct species label. The benchmark contains 30 categories, each corresponding to a species (Yan et al., 13 Aug 2025).

Component	Per-category count	Total
Target images	45 per species	1,350 images
Reference images	$\sim 100$ per species	3,000 images

For the target set, each species contributes 45 images, split into 40 train and 5 test. The resulting train set comprises $30 \times 40 = 1{,}200$ HQ-LQ pairs, while the validation/test set comprises $30 \times 5 = 150$ HQ-LQ pairs. The same reference pool is used, held out from target images. This structure makes the benchmark category-organized rather than pair-organized, which is central to its retrieval-augmented character.

LQ-HQ pair generation follows the same degradation pipeline as Real-ESRGAN, specifically random blur, noise, JPEG compression, and downsampling. For each HQ target image, its LQ version is generated through that pipeline. During training, HQ targets and references are cropped into $768 \times 768$ patches. The protocol then pre-retrieves top-5 references by cosine similarity in embedding space per HQ and randomly samples one per iteration; for each sampled reference, $512 \times 512$ patches are randomly cropped from HQ and reference. Augmentation consists of random horizontal flipping and rotation. These design choices indicate that retrieval is integrated into the training protocol rather than appended only at inference time.

3. Retrieval formulation and reference database organization

The retrieval function is defined as top-$K$ nearest neighbors under cosine similarity in an embedding space $f(\cdot)$:

$$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$0

with

$$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$1

and

$$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$2

The semantic reference retriever uses a pretrained DINOv2 vision transformer as the feature extractor $$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$3. Offline, all reference images $$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$4 are encoded and their embeddings are stored in a vector database, Qdrant. Online, the LQ input is encoded to $$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$5, cosine similarity is computed against stored embeddings, and the top-$$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$6 references are retrieved. The selected retrieval set is denoted $$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$7, with $$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$8 at test time and $$\mathrm{RASR}(x_{LQ}) = \mathcal{G}(x_{LQ}, R(x_{LQ})),$$9 during train pre-selection (Yan et al., 13 Aug 2025).

This arrangement is significant because it makes retrieval quality a measurable factor in restoration quality. The benchmark therefore does not only evaluate reconstruction from degraded inputs; it also evaluates the quality of semantic indexing and nearest-neighbor selection under realistic database search.

4. Baseline system: RASRNet

RASRNet is the baseline model proposed alongside RASR-Flickr30. Its architecture is summarized as RIRR (retriever) plus RefDSR (generator). RefDSR is built on a pretrained latent diffusion SISR model with a frozen VAE encoder $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$0, decoder $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$1, and UNet $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$2. Reference conditioning is injected through a ControlNet branch $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$3 with LoRA adapters (Yan et al., 13 Aug 2025).

Conditioning combines text and visual reference signals. Separate captions $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$4 and $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$5 are extracted for the LQ input and each retrieved reference, then concatenated as $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$6. Reference features are defined as

$R(x_{LQ}) \subseteq \mathcal{D}_{ref}$7

Fusion occurs in the UNet decoder at the first three blocks $R(x_{LQ}) \subseteq \mathcal{D}_{ref}$8. During training,

$R(x_{LQ}) \subseteq \mathcal{D}_{ref}$9

and during inference, balanced fusion is used:

$\sim 100$0

The final decoder block uses cross-attention to $\sim 100$1. The denoising objective is given as

$\sim 100$2

Additional loss terms are an MSE loss,

$\sim 100$3

a perceptual LPIPS loss,

$\sim 100$4

a GAN loss with a DINO-based discriminator,

$\sim 100$5

and a Gram loss,

$\sim 100$6

where $\sim 100$7 is the Gram matrix of VGG-16 layer $\sim 100$8. The total training loss is

$\sim 100$9

The model therefore combines latent diffusion restoration, semantic retrieval, and explicit reference-conditioned control. A plausible implication is that the benchmark is intended not merely for conventional SR scoring, but for studying how retrieval quality, conditioning strength, and perceptual losses interact.

5. Quantitative results and ablation structure

On the RASR-Flickr30 test set, the reported comparison is against OSEDiff as baseline. RASRNet improves PSNR, SSIM, MUSIQ, and CLIPIQA, while reducing LPIPS, DISTS, FID, and NIQE (Yan et al., 13 Aug 2025).

Metric	OSEDiff	RASRNet
PSNR	23.29	23.67
SSIM	0.5624	0.5687
LPIPS	0.3404	0.3273
FID	65.59	56.83

The paper summarizes these gains as $30 \times 40 = 1{,}200$0 dB PSNR and $30 \times 40 = 1{,}200$1 LPIPS, with more realistic textures. Qualitative examples include a puffin case in which a retrieved puffin eye image is used to reconstruct realistic iris patterns, and a peacock case in which intricate feather patterns are injected more effectively than in diffusion-only methods.

Ablation studies examine prompt conditioning, loss composition, and retrieval encoder choice. For text prompt conditioning with $30 \times 40 = 1{,}200$2 retrieval, the combination REF + LR yields PSNR $30 \times 40 = 1{,}200$3, LPIPS $30 \times 40 = 1{,}200$4, FID $30 \times 40 = 1{,}200$5, and CLIPIQA $30 \times 40 = 1{,}200$6, outperforming the no-prompt, LR-only, and REF-only variants. For loss components, the joint use of Gram and GAN losses yields the best CLIPIQA, with the configuration $30 \times 40 = 1{,}200$7, $30 \times 40 = 1{,}200$8 giving PSNR $30 \times 40 = 1{,}200$9, FID $30 \times 5 = 150$0, NIQE $30 \times 5 = 150$1, and CLIPIQA $30 \times 5 = 150$2. For retrieval encoder comparison, random retrieval is clearly worse than VGG16, CLIP, and DINOv2, while DINOv2 is the selected retriever in the final system. Additional evaluation on the WR-SR test set reports RASRNet at PSNR $30 \times 5 = 150$3, SSIM $30 \times 5 = 150$4, LPIPS $30 \times 5 = 150$5, FID $30 \times 5 = 150$6, and NIQE $30 \times 5 = 150$7.

These results position retrieval as a substantive source of performance gain rather than a cosmetic add-on. The ablations also show that the benchmark supports controlled study of retrieval backbones and conditioning pathways under a fixed restoration architecture.

6. Practical implications, limits, and relation to other Flickr-derived datasets

The practical implications identified for RASR-Flickr30 are open-world retrieval, scalability, and flexibility. Per-category reference pools eliminate the need for manually paired references. New species or object categories can be added by collecting a small reference database, without retraining the retriever. The same database can support multiple LQ inputs across different scenes, including zoos, museums, and wildlife photography (Yan et al., 13 Aug 2025).

The limitations are also explicit. Retrieval latency is reported as $30 \times 5 = 150$8 per query on $30 \times 5 = 150$9 images in Qdrant. Single-step diffusion plus retrieval overhead is described as promising for real-time use, but end-to-end latency remains a barrier for mobile deployment. The move beyond 30 species to general “in-the-wild” scenes is said to require semantic clustering and dynamic database updates. Fine details, particularly very small text or subtle patterns, remain difficult, motivating possible future work on patch-level retrieval or multi-scale reference fusion.

The naming of RASR-Flickr30 may invite comparison with other Flickr-derived datasets, but those resources address different problems. Flickr30K, analyzed in the context of stereotype-driven descriptions and unwarranted inferences, is a crowdsourced caption dataset rather than an image restoration benchmark (Miltenburg, 2016). The extended SPID 2018 dataset used for predicting popularity over 30 days is organized around daily view-count sequences, social metadata, and visual features for approximately 20 000 Flickr images (Dutta et al., 2021). This suggests that “Flickr” in RASR-Flickr30 primarily indicates image provenance, while “30” denotes the benchmark’s 30 animal species rather than a shared annotation protocol or prediction task.

Within super-resolution research, the central contribution of RASR-Flickr30 is therefore infrastructural as much as algorithmic: it establishes a benchmark in which retrieval from a category-organized reference database is part of the formal problem definition. That makes it suitable for evaluating whether retrieval augmentation can bridge the gap between academic RefSR research and real-world applicability, which is the core claim advanced by the benchmark and its baseline system (Yan et al., 13 Aug 2025).