Research map / Image generation / Super-resolution

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

NeurIPS 2026 · Image generation › Super-resolution · approach: Perceptual and real-world SR · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LucidNFT, a multi-reward DiffusionNFT-based RL framework for flow-matching real-world super-resolution that uses a learned LR-referenced faithfulness evaluator (LucidConsistency), decoupled per-objective reward normalization, and a 20K real-degradation dataset (LucidLR) to improve perceptual quality while keeping LR consistency.

Problem

Generative real-world super-resolution produces sharp outputs that can hallucinate content unsupported by the low-resolution input, and no reliable LR-referenced faithfulness reward exists for preference-based RL alignment; scalarizing multiple rewards before normalization also compresses advantage

Approach

Train an LR-referenced consistency evaluator with LoRA on a frozen Qwen3-VL-Embedding using degradation-invariant content pools and original-inpainted hard negatives, and normalize each reward objective within each LR-conditioned rollout group before fusing them into the DiffusionNFT reward weight.

Key results

• LucidFlux(+LucidNFT) on RealLQ250 raises UniPercept IQA from 70.93 to 73.48 and NIQE drops from 3.74 to 3.25 relative to LucidFlux.
• LucidConsistency human-agreement on RealSR reaches 0.690 Agreement and 0.558 Recall@1 versus 0.643 and 0.465 for frozen Qwen3-VL-Embedding-8B.
• DiT4SR(+LucidNFT) improves most perceptual metrics on RealLQ250 while LucidConsistency rises from 0.9052 to 0.9172.

Surprising findings

• Perceptual metrics score substantially worse than the semantic embedding baseline in the human LR-faithfulness study, suggesting visual quality alone cannot identify hallucinations.
• The paper reports that as models get better at suppressing degradation, realism and LR-faithfulness begin to diverge, making hallucination a more critical failure mode.

Limitations

• LucidConsistency is part of the training reward, so its gains cannot serve as standalone evidence of faithfulness; the human study is small (RealSR, five annotators).
• Perceptual gains on some metrics are modest (e.g., CLIP-IQA on DRealSR for LucidFlux(+DPO) slightly decreases), and LucidConsistency is not uniformly improved across all backbones/benchmarks.
• DP2O-SR could not be compared directly because its training code is not released; a Diffusion-DPO re-implementation is used instead.

Experimental setup

Datasets: LucidLR LSDIR RAISE SAGI RealLQ250 DRealSR RealSR

Benchmarks: RealLQ250 DRealSR RealSR

Metrics: CLIP-IQA+ Q-Align MUSIQ MANIQA CLIP-IQA NIQE UniPercept IQA VisualQuality-R1 LucidConsistency Agreement

Baselines: DiffBIRv2 SeeSR DreamClear SUPIR DiT4SR LucidFlux Diffusion-DPO

Builds on: DiffusionNFT Diffusion-DPO Qwen3-VL-Embedding Real-ESRGAN LucidFlux

Closest papers by meaning

• DP²O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution (NeurIPS 2025 · similarity 0.92)
• Chain-of-Zoom: Extreme Super-Resolution via Scale Autoregression and Preference Alignment (NeurIPS 2025 · similarity 0.89)
• When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.87)
• SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution (CVPR 2024 · similarity 0.87)
• Auto-Encoded Supervision for Perceptual Image Super-Resolution (CVPR 2025 · similarity 0.87)
• Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · similarity 0.87)