GuideSR: Rethinking Guidance for One-Step High-Fidelity Diffusion-Based Super-Resolution

TL;DR

GuideSR tackles fidelity bottlenecks in single-step diffusion-based SR by separating structural guidance from generative refinement. It introduces a dual-branch architecture: a full-resolution Guidance Branch that preserves structure, and a latent Diffusion Branch that enhances perceptual quality, with FRBs, IGN, and LoRA tuning enabling efficient fusion. The paper reports state-of-the-art results across real-world SR benchmarks, achieving improvements in PSNR, SSIM, LPIPS, DISTS, and FID, including up to 1.39 dB PSNR gain on challenging DRealSR data, while maintaining one-step inference. This work demonstrates that faithful, high-quality restoration can be achieved with efficient diffusion priors and points to future directions in efficiency optimization and video restoration.

Abstract

In this paper, we propose GuideSR, a novel single-step diffusion-based image super-resolution (SR) model specifically designed to enhance image fidelity. Existing diffusion-based SR approaches typically adapt pre-trained generative models to image restoration tasks by adding extra conditioning on a VAE-downsampled representation of the degraded input, which often compromises structural fidelity. GuideSR addresses this limitation by introducing a dual-branch architecture comprising: (1) a Guidance Branch that preserves high-fidelity structures from the original-resolution degraded input, and (2) a Diffusion Branch, which a pre-trained latent diffusion model to enhance perceptual quality. Unlike conventional conditioning mechanisms, our Guidance Branch features a tailored structure for image restoration tasks, combining Full Resolution Blocks (FRBs) with channel attention and an Image Guidance Network (IGN) with guided attention. By embedding detailed structural information directly into the restoration pipeline, GuideSR produces sharper and more visually consistent results. Extensive experiments on benchmark datasets demonstrate that GuideSR achieves state-of-the-art performance while maintaining the low computational cost of single-step approaches, with up to 1.39dB PSNR gain on challenging real-world datasets. Our approach consistently outperforms existing methods across various reference-based metrics including PSNR, SSIM, LPIPS, DISTS and FID, further representing a practical advancement for real-world image restoration.

Figures (5)

• Figure 1: Architecture comparison of diffusion-based super-resolution approaches. (a) Standard diffusion process (e.g., OSEDiff osediff) processes latent representations directly; (b) Controller-based methods (e.g., DiffBIR diffbir, SeeSR seesr, StableSR stablesr) employ conditional mechanisms to guide the diffusion process; (c) Our proposed GuideSR introduces a dual-branch architecture with full-resolution feature guidance, preserving high-frequency details from the original input while leveraging the generative capabilities of diffusion models. This approach addresses the key limitation of existing methods the loss of structural fidelity due to VAE encoding of degraded inputs.
• Figure 2: Overview of GuideSR architecture. Our method introduces a dual-branch architecture where the Guidance Branch (top) processes full-resolution input to preserve high-frequency details, while the Diffusion Branch (bottom) operates in the latent space for enhanced perceptual quality. The Guidance Branch applies a series of Full Resolution Blocks (FRBs) and an Image Guidance Network (IGN) to output a refined image $R_2$, which are enabled by the channel attention mechanisms and feature refinement operations detailed in the bottom. The Diffusion Branch employs a LoRA-finetuned diffusion model to produce the final output $R_1$, where the features from the Guidance Branch are adaptively refined and integrated into the denoise U-Net. Both branches are supervised through discriminators with shared weights during training.
• Figure 3: Multi-metric performance visualization. We visualize the performance of different SR methods, displaying reference-based metrics (PSNR, SSIM, LPIPS, DISTS, FID) on the left and no-reference metrics (CLIPIQA, MANIQA, MUSIQ) on the right. Our GuideSR model prioritizes fidelity, consistently achieving the best reference-based metrics across all datasets. It does not lead in no-reference metrics due to the perception-distortion tradeoff blau2018perception. Despite this, GuideSR consistently covers the largest area across all datasets, demonstrating a superior balance across various quality aspects.
• Figure 4: Pixel-Space Fidelity (MSE) and Feature-Space Fidelity (LPIPS). Enhancing both MSE and LPIPS is generally challenging because boosting generative capabilities often increases feature-space fidelity and enhances the realism of restored images, but usually at the cost of reduced pixel-space fidelity. GuideSR achieves the highest scores in both MSE and LPIPS among all methods, demonstrating its ability to maintain fidelity in both pixel and feature spaces.
• Figure 5: Visual Comparison on Real-World Images from RealSR realsr and DRealSR drealsr. (Top) GuideSR accurately restores detailed features such as the text shape and the reflections on metal surfaces. (Middle) GuideSR correctly reconstructs the geometry of the concrete blocks, while OSEDiff introduces incorrect textures and a color shift. (Bottom) GuideSR faithfully restores the text, particularly the inverted "a" letter, whereas PASD generates authentic but slightly blurred text.