Organizations: Institute of Automation, Chinese Academy of Sciences · School of Artificial Intelligence, University of Chinese Academy of Sciences · OPPO AI Center, OPPO Inc.
Recent advances in generative super-resolution (SR) have greatly improved visual realism, yet existing evaluation and optimization frameworks remain misaligned with human perception. Full-Reference and No-Reference metrics often fail to reflect perceptual preference, either penalizing semantically plausible details due to pixel misalignment or favoring visually sharp but inconsistent artifacts. Moreover, most SR methods rely on ground-truth (GT)-dependent distribution matching, which does not necessarily correspond to human judgments. In this work, we propose RefReward-SR, a low-resolution (LR) reference-aware reward model for preference-aligned SR. Instead of relying on GT supervision or NR evaluation, RefReward-SR assesses high-resolution (HR) reconstructions conditioned on their LR inputs, treating the LR image as a semantic anchor. Leveraging the visual-linguistic priors of a Multimodal Large Language Models (MLLM), it evaluates semantic consistency and plausibility in a reasoning-aware manner. To support this paradigm, we construct RefSR-18K, the first large-scale LR-conditioned preference dataset for SR, providing pairwise rankings based on LR-HR consistency and HR naturalness. We fine-tune the MLLM with Group Relative Policy Optimization (GRPO) using LR-conditioned ranking rewards, and further integrate GRPO into SR model training with RefReward-SR as the core reward signal for preference-aligned generation. Extensive experiments show that our framework achieves substantially better alignment with human judgments, producing reconstructions that preserve semantic consistency while enhancing perceptual plausibility and visual naturalness. Code, models, and datasets will be released upon paper acceptance.
Figures & tables
Figure 1: Overview of our proposed evaluation paradigm. (a) HR A preserves structures, while HR B shows severe distortions. (b) RefReward-SR uses an MLLM for interpretable, LR-conditioned semantic scoring. (c) Existing metrics wrongly favor the distorted HR B, whereas ours aligns with human perception. (d) RefReward-SR uniquely ensures semantic fidelity and preference alignment without GT dependency. The asterisk(*) indicates that FR metrics evaluate high-level semantics only passively via low-level correspondence.
Figure 2: The proposed RefReward-SR framework. (a) MLLM Fine-tuning : Fine-tuning via GRPO with format and LR-conditioned rank rewards from RefSR-18K for human preference alignment. (b) Global-Local Crop Scoring : Extracting representative local crops via RAM and Grounding DINO, then fusing multi-scale MLLM scores via area-weighted averaging for comprehensive evaluation.
Method
In-Domain
Out-of-Domain
Agreement
Recall@1
Filter@1
Agreement
Recall@1
Filter@1
Human Agreement
Annotators
84.7±1.4
-
-
81.8±1.3
-
-
FR Metrics
PSNR
41.8±1.6
45.0
9.5
35.5±2.3
25.5
8.5
SSIM [ 47 ]
51.8±2.0
51.0
25.5
43.3±1.8
26.0
19.0
Table 1: Quantitative results (in %) on human preference alignment. We report agreement with human annotators, as well as Recall@1 and Filter@1 on both in-domain and out-of-domain test sets. “Annotators” denotes the average agreement among human annotators. All ± values indicate the maximum deviation from the mean.
Datasets
Methods
PSNR ↑
SSIM ↑
LPIPS ↓
MUSIQ ↑
CLIPIQA ↑
NIMA ↑
TOPIQ-IAA ↑
Q-Align ↑
VQ-R1 ↑
StableSR
24.54
0.7004
0.3068
65.92
0.6325
4.7880
4.5738
3.2930
3.7670
DiffBIR
24.83
0.6501
0.3650
69.28
0.7053
4.9192
4.8973
3.7889
4.1030
SeeSR
25.14
0.7211
0.3007
69.82
0.6705
4.9196
4.8548
3.7190
3.9830
OSEDiff
25.15
0.7341
0.2921
69.09
0.6693
4.8953
4.7545
3.6962
4.0930
PiSA-SR
25.50
0.7418
0.2672
70.15
0.6698
4.8954
4.7431
3.6354
4.0180
RealSR
TSD-SR
23.40
0.6938
0.2823
71.26
0.7416
5.0439
4.8958
3.8444
4.0800
Table 2: Quantitative comparison of state-of-the-art Real-ISR methods. To evaluate the effectiveness of our alignment strategy, the best results among the base model (C-FLUX), its DP 2 O-fine-tuned variant (DP 2 O-FLUX), and our proposed method are highlighted with a light red background. The best and second best results across all methods are highlighted in red and blue , respectively.
Figure 3: The result of the user study win rate and average RefReward-SR score.
Figure 4: Qualitative comparison of state-of-the-art generative SR methods. Please zoom in for a better view. Additional results are provided in the Suppl. Mater. .
Figure 7
Methods
PSNR ↑
SSIM ↑
LPIPS ↓
MUSIQ ↑
CLIPIQA ↑
NIMA ↑
TOPIQ-IAA ↑
Q-Align ↑
VQ-R1 ↑
C-FLUX (Base)
24.42
0.6742
0.3376
68.70
0.6221
4.8799
4.6303
3.4943
3.9650
w/o RefReward-SR
23.89
0.6920
0.3211
73.48
0.7239
5.2534
5.0935
4.1656
4.1655
w/o LPIPS
21.98
0.5798
0.4026
74.55
0.7434
5.4195
5.2948
4.3521
4.3348
w/o DEQA
23.87
0.7099
0.2967
69.02
0.6153
5.0189
4.6552
3.5313
4.0188
Ours (Full)
24.68
0.7101
0.3202
71.70
0.7007
5.0567
4.9533
4.1596
4.2930
Table 4: Ablation study of reward components on the RealSR benchmark. The best and second best results are highlighted in red and blue , respectively.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: User interface of our annotation platform. The left panel shows the LR reference as a semantic anchor, while the right displays four HR candidates for annotators to rank from 1 (best) to 4 (worst).
Figure 7: Semantic categorization of the RefSR-18K dataset into five major groups.
Figure 8: Pairwise win-rate matrix among evaluated HR candidates.
Method
In-Domain
Out-of-Domain
Agreement
Recall@1
Filter@1
Agreement
Recall@1
Filter@1
RefReward-SR (Default)
85.0±2.4
84.5
78.0
80.2±2.2
77.0
73.0
RefReward-SR (No-Think)
84.9±2.1
84.5
79.0
79.8±2.3
77.0
73.0
Appendix
Table 5: Quantitative comparison (in %) between the default reasoning mode and the fast inference (No-Think) mode of RefReward-SR. All ± values indicate the maximum deviation from the mean across different annotator rankings.
Hyperparameter
Value
Base model
Qwen3-VL-8B-Instruct
Optimization strategy
DeepSpeed ZeRO Stage 3
Number of GPUs
8 × H20
Precision
BFloat16
Attention implementation
FlashAttention-2
Max prompt length
1,024
Appendix
Table 6: Training hyperparameters for RefReward-SR.
Hyperparameter
Value
Model & Architecture
Base model
C-FLUX (DP 2 O-SR [ 50 ] )
ControlNet
Frozen
VAE
Frozen
LoRA target modules
QKV & output projections of DiT Transformer
LoRA rank / alpha
32 / 64
Appendix
Table 7: Training hyperparameters for preference-aligned SR via GRPO.
Figure 9: Additional visual results of RefReward-SR evaluations, illustrating the visual reasoning trajectories and the predicted quality scores.
Figure 10: Visual comparisons of preference-aligned SR results on real-world datasets. Please zoom in for a better view.
Figure 11: Visual comparisons of preference-aligned SR results on the DIV2K-val dataset. Please zoom in for a better view.
Real-world image super-resolution (SR) is often designed with a single restoration objective, despite the current capacity of generative models to produce multiple high-quality reconstructions for the same input. In this paper, we argue that the best restoration strategy is subject to the specific restoration profile: a Faithful restoration prioritizes reference consistency, structure preservation, and hallucination suppression, whereas an Aesthetic restoration prioritizes visually pleasing and natural-looking details. We propose FoA-SR, a novel preference optimization approach to real-world SR based on profiles. To achieve this goal, FoA-SR starts with our supervised FLUX.2-based SR adapter (Flux2SR) trained with LR latent conditioning, flow matching, and image-space reconstruction losses for paired LR-to-HR image super-resolution. Following the development of the shared supervised super-resolution adapter, FoA-SR generates a shared stochastic candidate pool for each input image and ranks the same candidates using profile-specific Faithful and Aesthetic rewards to mine winner-loser pairs. These pairs are used to fine-tune separate LoRA adapters while keeping the base model frozen. Experiments on RealSR and DIV2K show that FoA-SR can steer the same SR adapter towards distinct restoration objectives: a Faithful adapter improves reference-consistent metrics while an Aesthetic adapter boosts metrics that measure perceptual quality without reference. Our candidate-pool analysis shows that Faithful and Aesthetic rewards frequently select different winners, and a Hybrid-LoRA ablation shows that collapsing both profiles into one reward yields an implicit compromise rather than explicit profile control.
Amjad Mahdi Alqarni, Peizhong Ju
Department of Computer Science University of Kentucky Lexington, KY, USA
Aligning generative real-world image super-resolution models with human visual preference is challenging due to the perception--fidelity trade-off and diverse, unknown degradations. Prior approaches rely on offline preference optimization and static metric aggregation, which are often non-interpretable and prone to pseudo-diversity under strong conditioning. We propose OARS, a process-aware online alignment framework built on COMPASS, a MLLM-based reward that evaluates the LR to SR transition by jointly modeling fidelity preservation and perceptual gain with an input-quality-adaptive trade-off. To train COMPASS, we curate COMPASS-20K spanning synthetic and real degradations, and introduce a three-stage perceptual annotation pipeline that yields calibrated, fine-grained training labels. Guided by COMPASS, OARS performs progressive online alignment from cold-start flow matching to full-reference and finally reference-free RL via shallow LoRA optimization for on-policy exploration. Extensive experiments and user studies demonstrate consistent perceptual improvements while maintaining fidelity, achieving state-of-the-art performance on Real-ISR benchmarks.
Shijie Zhao, Xuanyu Zhang, Bin Chen +8
ByteDance Inc. · Peking University · The Chinese University of Hong Kong
Generative-prior-based super-resolution (SR) methods leverage implicit knowledge learned from large-scale models to enhance low-resolution inputs. However, such priors are inherently stochastic and optimized for diverse generation, conflicting with the deterministic, high-fidelity reconstruction required for SR. It is also tied to the base model, costly to update, and prone to hallucinations as real-world data evolves. To address these limitations, we introduce WorldSR, a novel framework that augments SR with explicit, entity-aligned visual world knowledge. WorldSR automatically decomposes the input into semantic entities and performs entity-wise retrieval of structurally and semantically aligned images from a large-scale external corpus. For efficiency and reliability, we incorporate a memory mechanism to reuse previously retrieved results and a multi-step filtering process to refine candidate references. The selected references as world knowledge are then integrated through a novel multi-reference super-resolution model, which provides fine-grained constraints for detail synthesis and structural recovery. Experiments on real-world SR benchmarks show that WorldSR consistently improves structural fidelity and visual quality, outperforming existing generative super-resolution methods. Our results suggest that entity-level grounding in external visual world knowledge provides a simple and effective complement to model-bound generative priors.
Junyuan Deng, Xinyi Wu, Zhenyao Wu +3
Honor Device Co., Ltd. · Shenzhen University of Advanced Technology