OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution

Aligning generative real-world image super-resolution models with human visual preference is challenging due to the perception--fidelity trade-off and diverse, unknown degradations. Prior approaches rely on offline preference optimization and static metric aggregation, which are often non-interpretable and prone to pseudo-diversity under strong conditioning. We propose OARS, a process-aware online alignment framework built on COMPASS, a MLLM-based reward that evaluates the LR to SR transition by jointly modeling fidelity preservation and perceptual gain with an input-quality-adaptive trade-off. To train COMPASS, we curate COMPASS-20K spanning synthetic and real degradations, and introduce a three-stage perceptual annotation pipeline that yields calibrated, fine-grained training labels. Guided by COMPASS, OARS performs progressive online alignment from cold-start flow matching to full-reference and finally reference-free RL via shallow LoRA optimization for on-policy exploration. Extensive experiments and user studies demonstrate consistent perceptual improvements while maintaining fidelity, achieving state-of-the-art performance on Real-ISR benchmarks.

The effectiveness of super resolution (SR) models hinges on their ability to recover high frequency structure without introducing artifacts. Diffusion based approaches have recently advanced the state of the art in SR. However, most diffusion based SR pipelines operate purely in the spatial domain, which may yield high frequency details that are not well supported by the underlying low resolution evidence. On the other hand, unlike supervised SR models that may inject dataset specific textures, single image SR relies primarily on internal image statistics and can therefore be less prone to dataset-driven hallucinations; nevertheless, ambiguity in the LR observation can still lead to inconsistent high frequency details. To tackle this problem, we introduce BATDiff, an unsupervised Bivariate A trous Wavelet Diffusion model designed to provide structured cross scale guidance during the generative process. BATDiff employs an a Trous wavelet transform that constructs an undecimated multiscale representation in which high frequency components are progressively revealed while the full spatial resolution is preserved. As the core inference mechanism, BATDiff includes a bivariate cross scale module that models parent child dependencies between adjacent scales. It improves high frequency coherence and reduces mismatch artifacts in diffusion based SR. Experiments on standard benchmarks demonstrate that BATDiff produces sharper and more structurally consistent reconstructions than existing diffusion and non diffusion baselines, achieving improvements in fidelity and perceptual quality.

Diffusion models have recently demonstrated strong performance for image restoration tasks, including super-resolution. However, their large model size and iterative sampling procedures make them computationally expensive for practical deployment. In this work, we present TOC-SR, a framework for building efficient one-step super-resolution models by first discovering a compact diffusion backbone. Starting from a sixteen-channel latent diffusion model, we construct parameter-efficient surrogate blocks using feature-wise generative distillation and perform architecture discovery using epsilon-constrained Bayesian Optimization to minimize model complexity while preserving generative fidelity. The resulting compact diffusion backbone achieves a 6.6x reduction in parameters and a 2.8x reduction in GMACs compared to the expanded diffusion model. We then adapt this backbone for super-resolution and distill the diffusion process into a single-step generator. Experiments demonstrate that the proposed approach enables efficient super-resolution while maintaining strong reconstruction quality.