LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution".
Jane: Here's the gist of the paper. Generative real-world image super-resolution (Real-ISR) can synthesize visually convincing details from severely degraded low-resolution (LR) inputs, yet its 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 The authors propose LucidNFT, a multi-reward RL framework for flow-matching Real-ISR This framework introduces LucidConsistency, a degradation-invariant and hallucination-sensitive LR-referenced evaluator trained with content-consistent degradation pools and original–inpainted hard negatives Extensive experiments show that LucidNFT improves perceptual quality on strong flow-based Real-ISR baselines while generally maintaining LR-referenced consistency across diverse real-world scenarios.
Tom: Pruning redundant reasoning tokens The paper addresses the three coupled challenges hindering effective alignment in Real-ISR: (i) the lack of an LR-referenced faithfulness signal that is robust to degradation yet sensitive to localized hallucinations, (ii) a rollout-group optimization bottleneck where scalarizing heterogeneous rewards before normalization compresses objective-wise contrasts and weakens DiffusionNFT-style reward-weighted updates, and (iii) limited coverage of real degradations, which restricts rollout diversity and preference signal quality.
Jane: LucidNFT introduces three components to solve these issues: LucidConsistency, a degradation-invariant and hallucination-sensitive LR-referenced evaluator trained with content-consistent degradation pools and original–inpainted hard negatives, decoupled reward normalization, which preserves objective-wise reward 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.
Tom: How the training signal is built LucidConsistency is designed to be degradation-invariant and hallucination-sensitive by using content-consistent pools where same-content views under different degradation levels form positive pools, and original–inpainted pairs are used as controlled hard negatives.
Paper summary: Tom: Looking at the whole picture of LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution, the authors show that hallucination in current real world super resolution is capability dependent, meaning conservative models can stay faithful even if their perceptual quality isn't top tier.
Jane: That’s a crucial point because it suggests that as we build stronger models, the gap between how realistic they look and how faithful they are to the low resolution input starts to widen.
Lu: LucidNFT offers a path toward more reliable generative restoration by improving perceptual metrics while keeping those LR-inconsistent artifacts relatively low.
Meng: If you're deploying this stuff, it means you can push the quality up on perception without having to fight those structural hallucinations that break things when you put them in production.
Tom: So, for the end user, LucidNFT suggests a route toward generative restoration that is more dependable because it manages the trade-off between how sharp something looks and whether it actually matches what was there originally.
Conclusion: Tom: So we've been looking at LucidNFT, which is about using reinforcement learning to fix those blurry images that AI makes when it tries to super-resolve real world stuff from low resolution inputs.
Jane: Right, so the main idea is that this new method, LucidNFT, tackles a problem where the AI gets really creative and hallucinates details it shouldn't be making up.
Tom: It does this by introducing three things: first, something called LucidConsistency to check for hallucinations based on how we degrade real images; second, some way to handle rewards that doesn't mess up the signal you get from the low resolution input; and third, a big library of real degraded images for training.
Lu: Exactly. The authors are trying to build a system where the model learns what looks right based on actual degraded examples instead of just guessing based on some abstract score.
Meng: From my side, I'm looking at the numbers they show—they say this method improves perceptual quality compared to older flow-based methods, but it also keeps that LR reference consistent across different real world scenes.
Jane: That consistency is key for me. It means if you take a picture of a car and degrade it to low resolution, and then ask the AI to fix it, the fix looks good without totally changing what was actually in there.
Tom: And they found that their reward normalization technique keeps those perceptual rewards and faithfulness rewards from canceling each other out, which is a big deal for making sure the final output is both sharp *and* accurate.
Lu: The ablation study shows that training on this large collection of real degraded images really boosts both the perceptual quality and the overall performance on a specific benchmark called RealLQ250.
Jane: So what does this mean practically? It suggests a practical way to make generative restoration better for real world applications without just making things look pretty but being completely wrong about what's actually there.
Tom: It points toward models that are more reliable because they learn from the evidence of real degradation rather than just relying on abstract mathematical scores.
Lu: And the paper concludes that hallucination in these systems is tied to how strong the model is, meaning conservative models can still be faithful even if they aren't super sharp.
Jane: It’s a hopeful sign for deployment, though it also tells us where we need to focus next—making sure that realism doesn't completely drift away from what the original low resolution input was.
Hong Kong University of Science and Technology
cs.CV
Submitted: 2026-03-06
Updated: 2026-11-10
Code: https://github.com/W2GenAI-Lab/LucidNFT
Project page: https://w2genai-lab.github.io/LucidNFT
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 100/100
The gist: Generative real-world image super-resolution (Real-ISR) can synthesize visually convincing details from severely degraded low-resolution (LR) inputs…
Key concepts
- LucidConsistency
- This is a specialized evaluator trained on content-consistent degradation pools. It measures both global image embeddings and local token embeddings, specifically designed to be sensitive to hallucinations while remaining invariant to image degradation. It uses original-inpainted pairs as hard negatives to refine this sensitivity.
- Decoupled Reward Normalization
- This technique modifies how rewards are combined within a rollout group. Instead of scalarizing rewards before normalization, it computes individual reward contrasts for each component (perceptual and faithfulness) separately before fusing them. This prevents one type of reward from suppressing the other, preserving their distinct objective-wise signals.
- LucidLR
- This refers to a large-scale collection of real-world degraded images used for robust fine-tuning. By training the model on this diverse dataset, LucidNFT gains better coverage of real degradations, leading to more diverse rollout groups and higher quality preference signals during the RL process.
Terminology
Summary
Generative real-world image super-resolution (Real-ISR) can synthesize visually convincing details from severely degraded low-resolution (LR) inputs, yet its 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 The authors propose LucidNFT, a multi-reward RL framework for flow-matching Real-ISR This framework introduces LucidConsistency, a degradation-invariant and hallucination-sensitive LR-referenced evaluator trained with content-consistent degradation pools and original–inpainted hard negatives Extensive experiments show that LucidNFT improves perceptual quality on strong flow-based Real-ISR baselines while generally maintaining LR-referenced consistency across diverse real-world scenarios. p. 0 com/W2GenAI-Lab/LucidNFT
Abstract
Generative real-world image super-resolution (Real-ISR) can synthesize visually
convincing details from severely degraded low-resolution (LR) inputs, yet its
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. Preference-based reinforcement learning (RL) is a natural fit because
each LR input yields a rollout group of candidate restorations.Where the paper says it
Pruning redundant reasoning tokens
The paper addresses the three coupled challenges hindering effective alignment in Real-ISR: (i) the lack of an LR-referenced faithfulness signal that is robust to degradation yet sensitive to localized hallucinations, (ii) a rollout-group optimization bottleneck where scalarizing heterogeneous rewards before normalization compresses objective-wise contrasts and weakens DiffusionNFT-style reward-weighted updates, and (iii) limited coverage of real degradations, which restricts rollout diversity and preference signal quality. LucidNFT introduces three components to solve these issues: LucidConsistency, a degradation-invariant and hallucination-sensitive LR-referenced evaluator trained with content-consistent degradation pools and original–inpainted hard negatives, decoupled reward normalization, which preserves objective-wise reward 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. p. 0 We propose LucidNFT, a multi-reward RL framework for flow-matching
Real-ISR. LucidNFT 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 LRconditioned rollout group before fusion; and LucidLR, a large-scale collection
of real-world degraded images for robust RL fine-tuning.Where the paper says it
How the training signal is built
LucidConsistency is designed to be degradation-invariant and hallucination-sensitive by using content-consistent pools where same-content views under different degradation levels form positive pools, and original–inpainted pairs are used as controlled hard negatives. This evaluator extracts a global embedding eg(v) and native image-token embeddings Et(v), and the inference score C(xlr, xsr) combines global and local consistency terms. p. 4 (7)
Inference score. Given an LR input xlr and a generated SR output xsr, LucidConsistency combines
global and local consistency:
C(xlr, xsr) = αg s
eg(xlr), eg(xsr)
+ αl
1
N
X
N
i=1
s
e
i
t
(xlr), ei
t
(xsr)
, αg + αl = 1. We propose LucidNFT, a multi-reward RL framework for flow-matching
Real-ISR. LucidNFT 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 LRconditioned rollout group before fusion; and LucidLR, a large-scale collection
of real-world degraded images for robust RL fine-tuning. • Decoupled reward normalization, which preserves objective-wise reward contrasts within
each LR-conditioned rollout group before fusion. This prevents perceptual and faithfulness
rewards from suppressing each other and maintains discriminative reward weights for
forward fine-tuning.Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The effectiveness of LucidNFT is demonstrated through experiments on two flow-based Real-ISR backbones, showing that LucidNFT improves perceptual quality while generally maintaining LR-referenced consistency across diverse real-world scenarios. The ablation study confirms that training on LucidLR (E) achieves the best perceptual quality and the strongest overall RealLQ250 performance, while keeping LucidConsistency above the original LucidFlux baseline. p. 8 Replacing
scalar-first aggregation with decoupled reward normalization (D) further improves perceptual metrics
while maintaining high consistency, supporting the importance of preserving per-objective reward
contrast before fusion. Finally, training on LucidLR (E) achieves the best perceptual quality and
the strongest overall RealLQ250 performance, while keeping LucidConsistency above the original
LucidFlux baseline. As shown in Appendix Fig. 10, LucidNFT
consistently yields larger advantage gaps and higher DAGC than scalar-first reward aggregation
under the same DiffusionNFT objective, indicating that decoupled normalization better preserves
perceptual–faithfulness trade-offs for reward-weighted fine-tuning.Where the paper says it
Where the paper says it
Discussion
The study concludes that hallucination in current Real-ISR is capability-dependent, suggesting that conservative models can remain faithful despite low perceptual quality, but realism and LR-faithfulness begin to diverge as models become stronger. The results suggest that LucidNFT provides a practical route toward more reliable generative restoration by improving perceptual metrics without visibly increasing LR-inconsistent artifacts. p. 9 two flow-based Real-ISR backbones show improved perceptual quality while generally maintaining
LR-referenced consistency, suggesting a practical route toward more reliable generative restoration. Discussion.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improvement of Real-ISR models by incorporating LucidConsistency allows them to synthesize details that are both visually convincing and faithful to LR evidence, preventing semantic or structural hallucinations by learning a
degradation-invariant and hallucination-sensitive LR-referenced evaluator.
-
Improvement of reward optimization through Decoupled Reward Normalization ensures that perceptual quality and LR faithfulness rewards do not suppress each other, preserving
objective-wise contrasts within each LR-conditioned rollout group before fusion,
which leads to better preference guidance during forward fine-tuning. -
Improvement of training diversity via LucidLR expands the set of real-world degraded inputs, providing
diverse LR inputs for informative stochastic rollouts and stable preference learning
necessary for robust RL fine-tuning across heterogeneous degradation patterns.
Abstract
Generative real-world image super-resolution (Real-ISR) can synthesize visually convincing details from severely degraded low-resolution (LR) inputs, but stochastic sampling can introduce semantic or structural hallucinations unsupported by the LR evidence. Sampling multiple outputs per LR input exposes trade-offs between perceptual quality and faithfulness that reinforcement learning (RL) can optimize beyond paired supervision. Effective alignment is limited by the lack of a degradation-robust LR-referenced faithfulness signal, the loss of per-objective reward contrasts when rewards are combined before normalization in Diffusion Negative-aware FineTuning (DiffusionNFT), and limited coverage of real degradations. We propose LucidNFT, a multi-reward RL framework for flow-matching Real-ISR. LucidNFT introduces LucidConsistency, a degradation-invariant LR-referenced evaluator trained with content-consistent degradation pools and hard negatives created by inpainting to detect localized hallucinations. Its decoupled reward normalization strategy normalizes each reward within each LR-conditioned rollout group before fusion, preserving per-objective contrasts across candidates. LucidLR provides a large-scale collection of real-world degraded images for RL fine-tuning. Experiments show that LucidNFT improves perceptual quality on strong flow-based Real-ISR baselines while generally maintaining LR-referenced consistency across diverse real-world scenarios. In a reward-independent human study, LucidNFT outputs were judged more faithful to the LR inputs in 72.5 percent of comparisons.
Sources
- Training Diffusion Models with Reinforcement Learning
- DSPO: Direct Semantic Preference Optimization for Real-World Image Super-Resolution
- LucidFlux: Caption-Free Photo-Realistic Image Restoration via a Large-Scale Diffusion Transformer
- LPNSR: Learnable Noise Prediction for Diffusion-Based Image Super-Resolution
- Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking
- Flow Matching for Generative Modeling
- Flow-GRPO: Training Flow Matching Models via Online RL
- Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
- FinPercep-RM: A Fine-grained Reward Model and Co-evolutionary Curriculum for RL-based Real-world Super-Resolution
- Decoupled Weight Decay Regularization
- SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis
- RealSR-R1: Reinforcement Learning for Real-World Image Super-Resolution with Vision-Language Chain-of-Thought
- XPSR: Cross-modal Priors for Diffusion-based Image Super-Resolution
- Proximal Policy Optimization Algorithms
- RefReward-SR: LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution
- Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels
- DP$^2$O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution
- VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank
- DanceGRPO: Unleashing GRPO on Visual Generation
- GDPO-SR: Group Direct Preference Optimization for One-Step Generative Image Super-Resolution
Related papers
- Efficient All-in-One Weather Restoration using Spectral Harmonization
- Peak-End-Net: A Peak-End Rule Inspired Framework for Generalizable Video Aesthetic Assessment
- Sensor2Sensor: Cross-Embodiment Sensor Conversion for Autonomous Driving
- WithEveryone: Unified Planning and Identity Grounding for Group Image Generation
- SciFigure2Code: An AI-Reconstructed Benchmark for Scientific Figure-to-Code
- RepFusion: Leveraging Multimodal Priors for Denoising in Representation Space