LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution

NeurIPS 2026

Rollout-group multi-reward fine-tuning for flow-based Real-ISR with LR-referenced faithfulness.

Song Fei1, †, Tian Ye1, †, Sixiang Chen1, Zhaohu Xing1, Jianyu Lai1, Lei Zhu1, 2, *

1 The Hong Kong University of Science and Technology (Guangzhou), 2 The Hong Kong University of Science and Technology

† Equal contribution   * Corresponding author

Abstract

Generative real-world image super-resolution can synthesize visually convincing details from severely degraded low-resolution inputs, yet stochastic sampling makes a critical failure mode hard to avoid: outputs may look sharp but be unfaithful to the LR evidence, exhibiting semantic or structural hallucinations.

LucidNFT is a multi-reward RL framework for flow-matching Real-ISR. It introduces LucidConsistency, a degradation-invariant and hallucination-sensitive LR-referenced evaluator trained with content-consistent degradation pools and original-inpainted hard negatives; a decoupled reward normalization strategy that preserves objective-wise contrasts within each LR-conditioned rollout group before fusion; and LucidLR, a large-scale collection of real-world degraded images for robust RL fine-tuning.

Across two flow-based Real-ISR backbones, LucidNFT improves perceptual quality while generally maintaining LR-referenced faithfulness. In a separate human study, experts preferred its outputs as more faithful to the LR input in 72.5% of 200 comparisons.

LR faithfulness is missing

Without HR references, no-reference perceptual metrics can reward sharp but unsupported details. Real-ISR needs an LR-referenced signal that is robust to degradations and sensitive to hallucination.

Rollout groups need contrast

Preference learning compares multiple stochastic restorations conditioned on the same LR image. Scalarizing heterogeneous rewards before normalization can collapse perceptual-faithfulness distinctions.

Real degradations need scale

RL alignment benefits from diverse LR inputs that induce informative rollout variation. Small benchmark datasets and synthetic pipelines limit degradation coverage.

Method

LucidNFT overview: LucidConsistency training and scoring, followed by decoupled reward normalization for flow-based super-resolution
LucidNFT overview. LucidConsistency learns from content-consistent degraded views and original-inpainted hard negatives, then scores an SR output against its LR input. During RL fine-tuning, the perceptual and faithfulness rewards are normalized separately within each rollout group before their advantages are fused.

LucidConsistency

A Qwen3-VL-Embedding-8B backbone with trainable LoRA adapters learns global and native-token representations through pool-based contrastive losses. At inference, it combines global and local LR-SR consistency into an LR-referenced score.

Decoupled reward normalization

Each reward dimension is normalized within the same LR-conditioned rollout group before fusion. The fused advantage is stabilized at batch level and mapped to the bounded DiffusionNFT reward weight.

LucidNFT fine-tuning

LucidNFT uses UniPercept IQA as the perceptual reward and LucidConsistency as the LR-faithfulness reward. Fine-tuning uses LoRA rank 32, 24 LR groups of 12 rollouts per step, and LucidLR as the real-world LR source.

Camera-ready analysis of reward spread, reward influence, and the resolution of the training signal under scalar-first versus decoupled normalization
Why decoupled normalization matters. Across 16 LR inputs with 12 LucidFlux rollouts each, the perceptual reward varies about 2.9 times more within a group than the faithfulness reward. Scalar-first fusion follows the larger-scale reward and compresses candidate differences; decoupled normalization gives both rewards comparable influence and resolves more advantage levels.
Advantage gaps with scalar-first and decoupled reward normalization
Advantage-gap analysis on RealLQ250. Decoupled normalization preserves larger contrasts than scalar-first aggregation across rollout-group sizes; at 12 rollouts, the mean gap is 2.19 versus 0.72.

LucidLR Dataset

LucidLR contains 20,000 real-world low-quality photographs collected from Wikimedia Commons. A 21.7K-image raw pool was screened for unsafe content, very small images, and unavailable source pages. The released images have redistribution-compatible licenses with per-image source, author, and license metadata. Across the dataset, the paper identifies 13 semantic categories and 15 primary-degradation types.

LucidLR examples showing defocus blur, motion blur, noise, JPEG artifacts, low light, haze, color cast, and overexposure
Representative LucidLR inputs and their dominant real-world degradations. These low-quality photographs supply LR inputs for RL fine-tuning.
Real-world Real-ISR datasets. Degradation types exclude “other” and “none”; defocus is the share of images whose primary degradation is defocus blur.
Dataset Pairing Usage # Images Degr. Types Defocus
RealSRPairedBenchmark100482.0%
DRealSRPairedBenchmark93591.4%
RealLQ250UnpairedBenchmark2501043.2%
LucidLRUnpairedRL training20,0001526.5%
Semantic-category and primary-degradation distributions for LucidLR, RealLQ250, RealSR, and DRealSR
Camera-ready LucidLR statistics. Compared with RealLQ250, RealSR, and DRealSR under the same annotation taxonomy, LucidLR covers a broader mix of scene types and degradations, with defocus blur as the primary degradation in 26.5% of images.

Results

Experiments evaluate LucidNFT on two flow-based Real-ISR models, LucidFlux and DiT4SR, at 1024 × 1024 output resolution with 4× upscaling. On LucidFlux, the method improves all eight no-reference perceptual metrics across RealLQ250, DRealSR, and RealSR while generally maintaining LR-referenced consistency.

Independent human evaluation: For 40 held-out RealSR inputs, five expert annotators compared LucidFlux with LucidFlux(+LucidNFT) using LR faithfulness rather than sharpness or aesthetics. They preferred the LucidNFT outputs in 72.5% of 200 comparisons.

Training curves of LucidNFT and DPO fine-tuning
Training curves of LucidNFT and DPO fine-tuning. LucidNFT steadily improves perceptual rewards while keeping LucidConsistency stable.
Quantitative comparison on RealLQ250, DRealSR, and RealSR. Higher is better except NIQE.
Benchmark Metric DiffBIRv2 SeeSR DreamClear SUPIR DiT4SR DiT4SR(+LucidNFT) LucidFlux LucidFlux(+DPO) LucidFlux(+LucidNFT)
RealLQ250CLIP-IQA+ ↑0.69190.70340.68130.65320.70980.71240.72080.72280.7465
Q-Align ↑3.97554.14234.06474.13474.22704.23584.40524.44304.4855
MUSIQ ↑67.531370.375767.089965.813371.668272.173272.335172.450473.4475
MANIQA ↑0.49000.48950.44050.38260.46070.47190.52270.52580.5443
CLIP-IQA ↑0.71370.70630.69570.57670.71410.73550.68550.69170.7233
NIQE ↓5.11934.43833.87093.65913.55563.50073.74103.77853.2532
UniPercept IQA ↑65.476069.201568.846568.643073.074073.343070.930071.133073.4790
VisualQuality-R1 ↑4.34284.51184.44304.42654.61464.63044.54744.56444.6510
LucidConsistency ↑0.96090.94660.95780.95220.90520.91720.92370.92960.9345
DRealSRCLIP-IQA+ ↑0.64760.62580.44620.54940.65370.67570.65160.65300.6867
Q-Align ↑3.04873.27462.42143.47223.60083.66413.71413.74083.8423
MUSIQ ↑60.075961.322235.191254.928063.805165.191564.602564.560768.1545
MANIQA ↑0.49000.45050.26760.34830.44190.45720.46780.46690.5004
CLIP-IQA ↑0.67820.67600.43610.53100.67320.71110.66730.67130.7073
NIQE ↓6.48536.45037.01645.90925.70015.63295.07425.01434.1788
UniPercept IQA ↑46.229850.341434.247355.137158.129059.932859.903259.778263.7782
VisualQuality-R1 ↑3.47963.61162.56553.73493.96034.02393.99553.98284.1455
LucidConsistency ↑0.93320.92750.96070.89110.84380.85440.88130.88900.8879
RealSRCLIP-IQA+ ↑0.65430.67310.53310.56400.67530.68810.66690.66950.7151
Q-Align ↑3.31563.60733.00403.46823.71063.79593.87283.91473.9918
MUSIQ ↑61.775167.566049.476655.680767.982869.109267.896267.936270.5625
MANIQA ↑0.47450.50870.30920.34260.45330.46540.48890.49070.5284
CLIP-IQA ↑0.68060.69930.53900.48570.66310.69630.63590.64270.6936
NIQE ↓6.07005.45945.28735.28195.09124.83324.81344.68043.9526
UniPercept IQA ↑53.655058.053846.785056.606363.202564.842560.092560.477564.7588
VisualQuality-R1 ↑3.89284.06353.50283.78214.19534.24294.13764.15034.3389
LucidConsistency ↑0.95440.91380.94750.91410.83180.84980.88530.89320.8923
Human-aligned LR-faithfulness evaluation on RealSR.
EvaluatorCriterionAgreementRecall@1Filter@1
CLIP-IQAPerceptual Quality0.3910.1860.093
MUSIQPerceptual Quality0.3490.1160.093
Q-AlignPerceptual Quality0.3220.0930.047
UniPercept-IQAPerceptual Quality0.3220.0700.093
Qwen3-VL-Embedding-8BGeneric Semantics0.6430.4650.302
LucidConsistencyLR Faithfulness0.6900.5580.558
Controlled ablation on RealLQ250 using LucidFlux. The baseline row gives absolute scores; other rows are changes relative to that baseline.
MethodUniPercept ↑VQ-R1 ↑NIQE ↓LucidConsistency ↑
LucidFlux baseline70.9304.5473.7410.9237
(A) UniPercept IQA reward+0.608+0.054−0.227+0.0022
(B) (A) + frozen semantic reward+0.436+0.049−0.212+0.0061
(C) (B), semantic → LucidConsistency+0.284+0.046−0.194+0.0104
(D) (C), LSDIR → LucidLR+1.40+0.062−0.345+0.0114
(E) (C), scalar-first → decoupled+1.773+0.086−0.369+0.0119
(F) (E), LSDIR → LucidLR (full)+2.549+0.104−0.488+0.0108

Visual Results

Visual comparison on RealLQ250
Visual comparison on RealLQ250. LucidNFT variants recover more accurate text structures and finer local textures while preserving LR-supported semantics.
Visual comparison on RealLQ250
Visual comparison on RealLQ250. LucidFlux(+LucidNFT) improves local details while avoiding LR-inconsistent artifacts.

Citation

@inproceedings{fei2026lucidnft,
  title={LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution},
  author={Fei, Song and Ye, Tian and Chen, Sixiang and Xing, Zhaohu and Lai, Jianyu and Zhu, Lei},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026}
}

Contact Us

For questions or collaboration, contact sfei285@connect.hkust-gz.edu.cn, tye610@connect.hkust-gz.edu.cn, or leizhu@hkust-gz.edu.cn.