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

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

Song Fei^{1, †}, Tian Ye^{1, †}, Sixiang Chen¹, Zhaohu Xing¹, Jianyu Lai¹, Lei Zhu^{1, 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.

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

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, 12 rollouts per LR input, and LucidLR as the real-world LR source.

Advantage separability analysis on LucidFlux using RealLQ250. Decoupled normalization produces larger advantage gaps and more distinct advantage levels than scalar-first reward aggregation under the same DiffusionNFT objective.

LucidLR Dataset

LucidLR is a 20K-image real-world low-quality dataset collected from Wikimedia Commons through its official API. Images are gathered from public low-quality and blurred-image categories, filtered from an approximately 22K-image raw pool with NSFW classification, corrupted-file removal, and manual review.

Representative LucidLR samples with diverse real-world degradations, used as LR inputs for RL fine-tuning.

Comparison of representative real-world datasets used in Real-ISR.
Dataset	Pairing	Primary Usage	Type	# Images
RealSR	Paired	Testing / Benchmark	Real-captured	100
DRealSR	Paired	Testing / Benchmark	Real-captured	93
RealLQ250	Unpaired	Testing / Benchmark	Real-world	250
LucidLR	Unpaired	RL / Unsupervised Training	Real-world	20K

Results

Experiments evaluate LucidNFT on two flow-based Real-ISR models, LucidFlux and DiT4SR. All methods are evaluated at 1024 x 1024 output resolution with 4x upscaling. The paper reports eight no-reference quality metrics and LucidConsistency as an LR-referenced consistency score without HR ground truth.

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)
RealLQ250	CLIP-IQA+ ↑	0.6919	0.7034	0.6813	0.6532	0.7098	0.7124	0.7208	0.7228	0.7465
	Q-Align ↑	3.9755	4.1423	4.0647	4.1347	4.2270	4.2358	4.4052	4.4430	4.4855
	MUSIQ ↑	67.5313	70.3757	67.0899	65.8133	71.6682	72.1732	72.3351	72.4504	73.4475
	MANIQA ↑	0.4900	0.4895	0.4405	0.3826	0.4607	0.4719	0.5227	0.5258	0.5443
	CLIP-IQA ↑	0.7137	0.7063	0.6957	0.5767	0.7141	0.7355	0.6855	0.6917	0.7233
	NIQE ↓	5.1193	4.4383	3.8709	3.6591	3.5556	3.5007	3.7410	3.7785	3.2532
	UniPercept IQA ↑	65.4760	69.2015	68.8465	68.6430	73.0740	73.3430	70.9300	71.1330	73.4790
	VisualQuality-R1 ↑	4.3428	4.5118	4.4430	4.4265	4.6146	4.6304	4.5474	4.5644	4.6510
	LucidConsistency ↑	0.9609	0.9466	0.9578	0.9522	0.9052	0.9172	0.9237	0.9296	0.9345
DRealSR	CLIP-IQA+ ↑	0.6476	0.6258	0.4462	0.5494	0.6537	0.6757	0.6516	0.6530	0.6867
	Q-Align ↑	3.0487	3.2746	2.4214	3.4722	3.6008	3.6641	3.7141	3.7408	3.8423
	MUSIQ ↑	60.0759	61.3222	35.1912	54.9280	63.8051	65.1915	64.6025	64.5607	68.1545
	MANIQA ↑	0.4900	0.4505	0.2676	0.3483	0.4419	0.4572	0.4678	0.4669	0.5004
	CLIP-IQA ↑	0.6782	0.6760	0.4361	0.5310	0.6732	0.7111	0.6673	0.6713	0.7073
	NIQE ↓	6.4853	6.4503	7.0164	5.9092	5.7001	5.6329	5.0742	5.0143	4.1788
	UniPercept IQA ↑	46.2298	50.3414	34.2473	55.1371	58.1290	59.9328	59.9032	59.7782	63.7782
	VisualQuality-R1 ↑	3.4796	3.6116	2.5655	3.7349	3.9603	4.0239	3.9955	3.9828	4.1455
	LucidConsistency ↑	0.9332	0.9275	0.9607	0.8911	0.8438	0.8544	0.8813	0.8890	0.8879
RealSR	CLIP-IQA+ ↑	0.6543	0.6731	0.5331	0.5640	0.6753	0.6881	0.6669	0.6695	0.7151
	Q-Align ↑	3.3156	3.6073	3.0040	3.4682	3.7106	3.7959	3.8728	3.9147	3.9918
	MUSIQ ↑	61.7751	67.5660	49.4766	55.6807	67.9828	69.1092	67.8962	67.9362	70.5625
	MANIQA ↑	0.4745	0.5087	0.3092	0.3426	0.4533	0.4654	0.4889	0.4907	0.5284
	CLIP-IQA ↑	0.6806	0.6993	0.5390	0.4857	0.6631	0.6963	0.6359	0.6427	0.6936
	NIQE ↓	6.0700	5.4594	5.2873	5.2819	5.0912	4.8332	4.8134	4.6804	3.9526
	UniPercept IQA ↑	53.6550	58.0538	46.7850	56.6063	63.2025	64.8425	60.0925	60.4775	64.7588
	VisualQuality-R1 ↑	3.8928	4.0635	3.5028	3.7821	4.1953	4.2429	4.1376	4.1503	4.3389
	LucidConsistency ↑	0.9544	0.9138	0.9475	0.9141	0.8318	0.8498	0.8853	0.8932	0.8923

Human-aligned LR-faithfulness evaluation on RealSR.
Evaluator	Criterion	Agreement	Recall@1	Filter@1
CLIP-IQA	Perceptual Quality	0.391	0.186	0.093
MUSIQ	Perceptual Quality	0.349	0.116	0.093
Q-Align	Perceptual Quality	0.322	0.093	0.047
UniPercept-IQA	Perceptual Quality	0.322	0.070	0.093
Qwen3-VL-Embedding-8B	Generic Semantics	0.643	0.465	0.302
LucidConsistency	LR Faithfulness	0.690	0.558	0.558

Ablation study on RealLQ250 using LucidFlux.
Method	UniPercept ↑	VQ-R1 ↑	NIQE ↓	LucidConsistency ↑
LucidFlux baseline	70.930	4.547	3.741	0.9237
(A) IQA-only RL	71.538	4.601	3.514	0.9259
(B) + Frozen semantic reward	71.366	4.596	3.529	0.9298
(C) + LucidConsistency reward	71.214	4.593	3.547	0.9341
(D) + Decoupled norm.	72.703	4.633	3.372	0.9356
(E) + LucidLR	73.479	4.651	3.253	0.9345

Visual Results

Visual comparison on RealLQ250. LucidNFT variants recover more accurate text structures and finer local textures while preserving LR-supported semantics.

◀ ▶

LucidFlux

LucidFlux(+LucidNFT)

Example 016 LR

◀ ▶

LucidFlux

LucidFlux(+LucidNFT)

Example 137 LR

◀ ▶

LucidFlux

LucidFlux(+LucidNFT)

Example 223 LR

◀ ▶

LucidFlux

LucidFlux(+LucidNFT)

Example 182 LR

Citation

@article{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},
  journal={arXiv},
  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.