FGR-30k: Fine-Grained Super-Resolution Benchmark

• FGR-30k is a dedicated dataset containing 30,000 synthetic image-super resolution pairs with dense annotations to capture subtle local distortions.
• It uses diverse masking techniques, including block and semantic masks, and integrates pixel-wise and DINOv3 feature comparisons to create precise perceptual degradation maps.
• FGR-30k underpins the FinPercep-RM framework by enabling a co-evolutionary RLHF training curriculum that enhances both global fidelity and localized quality in super-resolution outputs.

The FGR-30k dataset is a dedicated large-scale benchmark for fine-grained perceptual evaluation in @@@@1@@@@ (ISR). It provides spatially dense annotations for local artifacts in synthesized super-resolved images, enabling the training of reward models that jointly assess global fidelity and dense local degradation. FGR-30k forms the foundation for the Fine-grained Perceptual Reward Model (FinPercep-RM), which significantly advances reinforcement learning with human feedback (RLHF) in visual generative tasks by suppressing reward hacking and improving both global and localized perceptual quality (Liu et al., 27 Dec 2025).

1. Construction and Composition

FGR-30k comprises 30,000 synthesized samples $\{I_{\rm syn}, M_{\rm gt}\}$, designed to capture diverse and subtle distortions characteristic of real-world super-resolution models. The construction process involves:

• Source Images ($I_{GT}$): High-quality natural photographs serve as ground-truth targets.
• Super-Resolved Outputs ($I_{SR}$): Samples of diffusion-based Real-ISR models applied to degraded low-resolution inputs.
• Distortion Synthesis: Artifacts are introduced via region swapping operations,

$I_{\rm syn} = M \odot I_{SR} + (1-M) \odot I_{GT}$

where $M$ is a mask (random block, free-form, or semantic from SAM) selecting spatial regions to inject artifacts, yielding locally realistic degradation.

Ground-Truth Perceptual Degradation Map ($M_{gt}$) provides dense spatial supervision, computed as a normalized, weighted fusion:

$$\begin{aligned} \text{Diff}_{\rm pixel} &= |I_{\rm syn} - I_{GT}| \ \text{Diff}_{\rm feat} &= 1 - \cos(f_{\rm syn}, f_{GT}) \quad \text{(features from DINOv3)} \ M_{gt} &= \text{Normalize}\left( \alpha\,\text{Diff}_{\rm pixel} + (1-\alpha)\,\text{Diff}_{\rm feat} \right) \end{aligned}$$

Artifact diversity is achieved via dozens of Real-ISR model outputs and multiple mask types (blocky versus semantic regions), ensuring broad coverage of realistic SR failure modes (Liu et al., 27 Dec 2025).

2. Supervised Learning Protocol

FGR-30k is primarily used to train reward models predicting:

• Global Quality Score ($S_{\rm fgc\text{-}global}$): A scalar estimating overall perceptual fidelity.
• Perceptual Degradation Map ($M_{\rm fg\text{-}pdm}$): A dense heatmap pinpointing artifact regions.

Given a candidate image $x$, the reward model outputs:

$$\left\{ S_{\rm fgc\text{-}global}, M_{\rm fg\text{-}pdm} \right\} = \text{RM}_\phi(x)$$

Training signals include:

• Dense map regression loss: $$\mathcal{L}_{\rm map} = \mathbb{E} \| M_{\rm fg\text{-}pdm} - M_{gt} \|_1$$.
• Triplet ranking over quality scores: $\mathcal{L}_{\rm rank}$ enforces correct ordering ($S_{GT} > S_{\rm fgc} > S_{SR}$).
• Score anchoring for scale stability: $\mathcal{L}_{\rm align}$ (L1 deviation from IQA baseline on reference images).

The total loss is a weighted sum:

$$\mathcal{L}_{\rm total} = \lambda_{\rm map} \,\mathcal{L}_{\rm map} + \lambda_{\rm rank} \,\mathcal{L}_{\rm rank} + \lambda_{\rm align} \,\mathcal{L}_{\rm align}$$

Training jointly calibrates both outputs, integrating local and global perceptual cues.

3. Dataset Statistics and Annotation Scheme

FGR-30k provides:

• Scale: 30,000 annotated samples, each with paired synthetic image and spatial degradation map.
• Source Diversity: Images processed through dozens of contemporary diffusion-based Real-ISR models.
• Masking Techniques: Use of block masks and semantic masks from Segment Anything Model (SAM), supporting both artificial and realistic artifact region selection.
• Feature Fusion: PDM ground-truth combines pixel-wise errors and feature-level anomalies (DINOv3), normalized to [0,1].

Component	Description	Quantity/Choices
Samples	($I_{\rm syn}$, $M_{gt}$) pairs	30,000
Source Images	High-quality photographs	Diverse
Models	Real-ISR diffusion-based	Dozens
Mask Types	Block, free-form, semantic (SAM)	2–3 strategies
Feature Backbone	DINOv3 for feature discrepancy	1

The density and diversity of labels facilitate spatially explicit perceptual learning and benchmark reward models for local defect sensitivity.

4. Application in Perceptual Reward Modeling

FGR-30k is the standard training set for FinPercep-RM, which utilizes an Encoder–Decoder architecture:

• Encoder: IQA backbone (e.g., CLIP-IQA) encodes multi-scale features.
• Decoder: Cross-layer upsampling constructs dense PDM heatmaps.
• Global Score Head: Modulates global score using PDM, incentivizing sensitivity to even subtle local artifacts.

The joint output enables reward models to provide dual supervision ("What" and "Where"), calibrated to penalize SR outputs with spatially concentrated distortions. This structure directly addresses limitations of previous scalar-output IQA models, which are prone to reward hacking due to insensitivity to spatial detail (Liu et al., 27 Dec 2025).

5. Role in RLHF Training Curricula

FGR-30k underpins a synchronized co-evolutionary curriculum for both the reward model and the generator (ISR model):

• Reward-Model Progressive Expansion: Starts with a frozen global IQA, incrementally augments model with decoder/adapters, and increases fine-grained sensitivity.
• Generator Curriculum Co-evolution: Generator policy is initially optimized on coarse reward, then gradually shifted to stricter fine-grained rewards as model complexity increases.

This easy-to-hard progression achieves stable training and robustness, as RLHF models transition smoothly from global to dense local supervision. Removing or weakening components leads to measurable drops in perceptual metrics, indicating the necessity of FGR-30k’s dense annotations for curriculum-based RL (Liu et al., 27 Dec 2025).

6. Experimental Impact and Benchmarks

Quantitative and user study evaluations demonstrate the efficacy of FGR-30k-trained FinPercep-RM across multiple 4× SR benchmarks:

• Metrics: Improvement across LPIPS (↓), MUSIQ (↑), MANIQA (↑), CLIP-IQA (↑), LIQE (↑).
• User Study: Realism and fidelity preferences align strongly with FinPercep-RM outputs (see Table 2).
• RLHF Strategy Independence: Gains hold across REFL, DPO, and GRPO algorithms—underlining that reward model quality, as imparted by FGR-30k, is orthogonal to RLHF optimization method.

A plausible implication is that future perceptual reward models will require datasets with explicit dense spatial annotation, such as FGR-30k, to effectively align generative models with human perceptual preferences and mitigate reward hacking.

7. Limitations and Prospects

Current PDM ground-truth in FGR-30k fuses only pixel-wise and DINOv3 feature-level discrepancies; expanding to richer multi-scale or semantic loss terms may further enhance defect localization. The co-evolutionary curriculum is manually staged, suggesting adaptive progression mechanisms could reduce convergence time. The FGR-30k paradigm may be extended to video super-resolution, denoising, deblurring, and multimodal reward learning scenarios (Liu et al., 27 Dec 2025).

FGR-30k establishes a new standard for fine-grained, spatially explicit reward model training in perceptual super-resolution research, enabling principled curriculum learning and robust alignment with qualitative human judgments.