MagnifiQ: Patch-aware Text Guided Progressive Upscaling for High-Resolution Image Restoration

Forum topic · 小凯 · 2026-08-18

Summary

MagnifiQ is an image restoration framework that progressively upscales and restores images from 1024x1024 to 4096x4096 resolution. Presented in arXiv paper 2508.08545 by Mahesh Reddy, Yashesh Savani, and Antoine Mercier, the method adapts a pre-trained text-to-image diffusion model (SDXL) for scalable high-resolution inference by replacing its self-attention layers with convolutional operations whose computational cost grows linearly with image resolution. A progressive upsampling strategy iteratively refines the image across multiple resolution stages rather than hallucinating the final 4K output at once, improving global consistency and reducing high-resolution artifacts. Patch-specific text prompts provide spatially localized semantic guidance to enhance local detail while controlling content drift. Experiments on synthetic and real-world degraded images show MagnifiQ outperforms prior diffusion-based restoration methods in perceptual quality and human preference, producing sharper textures and more consistent 4K results with a practical speed-quality trade-off.

Paper Overview

Field: Computer Vision (CV) Authors: Mahesh Reddy, Yashesh Savani, Antoine Mercier Published: 2026-08-17 arXiv: 2508.08545

Abstract (Translated)

High-resolution image restoration from degraded inputs is challenging because it must preserve global structural consistency while recovering fine-grained local details, especially at 4K resolution where direct diffusion-based restoration is computationally expensive and prone to repeated or inconsistent textures. This paper introduces MagnifiQ, a framework that progressively upscales and restores images across resolutions, e.g., from 1024x1024 to 4096x4096.

The approach leverages a pre-trained text-to-image diffusion model such as SDXL and adapts it for more scalable high-resolution inference by replacing its original self-attention layers with convolutional operations whose computational cost grows linearly with image resolution. A progressive upsampling strategy further iteratively restores and refines the image across multiple resolution stages instead of hallucinating the final 4K image directly, improving global consistency and reducing high-resolution artifacts. To enhance local detail while controlling content drift, MagnifiQ uses patch-specific text prompts to provide spatially localized semantic guidance during restoration.

Extensive experiments on synthetic and real-world degraded images show that MagnifiQ outperforms prior diffusion-based restoration methods in perceptual quality and human preference, generating sharper textures and more consistent 4K results while offering a practical speed-quality trade-off through its scalable backbone and progressive design.

Key Contributions

Scalable backbone: Replaces self-attention in SDXL with convolutions whose cost scales linearly with image resolution, enabling high-resolution diffusion inference
Progressive upsampling: Iteratively refines outputs across resolution stages (e.g., 1024x1024 → 4096x4096) for better global consistency
Patch-aware text guidance: Patch-specific prompts deliver spatially localized semantics, enhancing detail while limiting content drift
Results: Superior perceptual quality and human preference over prior diffusion-based restoration methods on synthetic and real-world degradations