MagnifiQ: Patch-aware Text Guided Progressive Upscaling for 4K Image Restoration

Forum topic · 小凯 · 2026-08-18

Summary

MagnifiQ is an image restoration framework that progressively upscales and restores images from resolutions such as 1024x1024 to 4096x4096, addressing the high computational cost and texture repetition problems of direct diffusion-based 4K restoration. The method builds on a pre-trained text-to-image diffusion model like SDXL, replacing its self-attention layers with convolutional operations whose cost grows linearly with image resolution for scalable inference. It introduces a progressive upscaling pipeline that restores and refines intermediate outputs across multiple resolution stages rather than hallucinating the final 4K image in one step, improving global consistency and reducing high-resolution artifacts. To sharpen local details while limiting content drift, MagnifiQ uses patch-specific text prompts that provide spatially localized semantic guidance during restoration. Experiments on synthetic and real-world degraded images show it 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 Authors: Mahesh Reddy, Yashesh Savani, Antoine Mercier Released: 2026-08-17 arXiv: 2508.08545

Abstract (Translation)

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.

MagnifiQ is 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.

The framework further proposes a progressive upscaling strategy: instead of directly hallucinating the final 4K image, it iteratively restores and refines the image across multiple resolution stages, improving global consistency and reducing high-resolution artifacts.

To enhance local details while controlling content drift, MagnifiQ uses patch-specific text prompts that provide spatially localized semantic guidance during the restoration process.

Original Abstract (excerpt)

> 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. In this work, we introduce MagnifiQ, an image restoration framework that progressively upscales and restores images across resolutions, e.g., from 1024x1024 to 4096x4096. Our 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. We further propose a progressive u...

Key Contributions

Scalable backbone: Self-attention replaced with convolutions for linear cost scaling with resolution
Progressive upscaling: Multi-stage refinement from 1024x1024 up to 4096x4096 instead of one-shot 4K generation
Patch-aware text guidance: Patch-specific prompts for spatially localized semantics that limit content drift
Results: Outperforms prior diffusion-based restoration methods in perceptual quality and human preference, with sharper textures and more consistent 4K output

Full paper: arXiv:2508.08545