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

원문 초록

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.