cs.CVNov 3, 2026

Wonder 4: A Mixture-of-Experts Recipe for Image Quality Enhancement

Authors: Yunan Wu, Bin Sun, Partha Acharjee, Xiaoyu Wang

Organizations: Topaz Labs Research

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

A real photograph is often blurry, grainy, compressed, scratched, and low-resolution at the same time. Each defect has its own severity and its own place in the image. The corrections for these defects conflict with each other. Exist- ing models describe the whole image with one description, so they cannot say where each correction belongs. To solve this problem, we present Wonder 4, a mixture-of-experts (MoE) model for image quality enhancement. It corrects several defects of a real photograph at the same time, each in the region where it occurs. Every block of the model holds a mixture of N = 13 experts. A router, trained with the model, sends every image patch to k = 5 of them. No degradation classifier, mask, prompt, detector, or task label is used. Another issue is that standard MoE training, taken from language models, pushes every expert toward equal use. Equal use applies every correction everywhere, even where it is not needed. We replace this rule with a bias that only sets a minimum share of use. Across all patches, an expert used less than this minimum gets higher router scores until it is selected again. An expert above the minimum is not changed, so no expert is forced to be used as often as the others. The bias needs no extra loss. The method adds no extra module and no extra processing step. We evaluate Wonder 4 on DIV2K and Flickr2K, against five published methods and three open-weight pipelines. It reaches the best LPIPS, DISTS, and NIQE on all three test settings (DIV2K x4, DIV2K x8, and Flickr2K x4).

Figures & tables

Figure 1: Wonder 4 at a glance. Top: a low-quality (LQ) input photograph, a map showing which expert each patch (a small region of the image) prefers in one block, and the enhanced output. Colors are expert indices. Bottom: the LQ image passes through LL blocks. In each block a router sends every patch to 5 of 13 experts (green cells). Because every patch chooses its own experts, a photograph with several defects is corrected in one pass with no task label and no degradation detector.
Figure 2: Wonder 4 inference path. The low-quality (LQ) input image yy is split into patches ( patchify ) and passed through a backbone of LL blocks of width dd , which produces the enhanced image x^\hat{x} . Every block carries its own router and its own bank of N=13N{=}13 experts (small cells). For each patch the router activates k=5k{=}5 of them (green, active ); the rest stay idle (grey). The active cells drawn differ between blocks because selection is made per patch and per block, not once per image.
Figure 3: Showcase: several defects mixed in real photographs. Each column shows the same crop from the input (top) and Wonder 4 (bottom). (a) Scanned print with dust and scratches over a blurred region. (b) Dust and scratches on wrapping paper. (c) Heavy grain in a low-light photograph. (d) Grain and softness on fine detail and lettering. (e) Grain on fur. (f) Compression blocking and softness on fur. (g) Blur and grain together. Every column carries at least two defect families at once, and Wonder 4 corrects them in a single pass with no detector, mask or task label. Best viewed at full size.
Figure 4: Per-patch routing, drawn on the image. Rows (a) and (b) are two photographs; the first tile of each row is the input (a is a square center crop; the grid size is given in patches, and the upscale factor is 1 ×\times for (a) and 2 ×\times for (b)). Each other tile is a map of one block, numbered from 00 at the network input. A patch is colored by the expert the router ranks first for it, aligned with the input. Early blocks are flat; late blocks follow the scene patch by patch. Each row has its own color key: only its four most-used experts are colored, all others are grey, and experts are identified only by their index.
 DIV2K wild ×4\times 4DIV2K bicubic ×8\times 8Flickr2K unknown ×4\times 4
MethodPSNR ↑\uparrowLPIPS ↓\downarrowDISTS ↓\downarrowNIQE ↓\downarrowPSNR ↑\uparrowLPIPS ↓\downarrowDISTS ↓\downarrowNIQE ↓\downarrowPSNR ↑\uparrowLPIPS ↓\downarrowDISTS ↓\downarrowNIQE ↓\downarrow
Real-ESRGAN [ 25 ]18.960.40470.21074.3623.430.39980.20165.8022.070.37230.18874.46
StableSR (Turbo) [ 26 ]19.090.44240.21866.0524.090.40130.19985.7822.040.39220.19505.83
SeeSR [ 27 ]18.910.36410.16123.7523.460.29270.13413.6622.020.33870.16493.76
SUPIR [ 28 ]19.070.36980.18493.8223.420.30140.14903.6622.640.39650.21944.75
OSEDiff [ 29 ]18.670.34790.16293.5923.100.28420.13553.6121.240.33910.16184.02
Table 1: Quantitative comparison on DIV2K validation (100 images, ×4\times 4 wild and ×8\times 8 bicubic) and Flickr2K ( 2 6502\,650 images, ×4\times 4 unknown degradation). Rows: five published enhancement methods (top), three open-weight pipelines (middle) and Wonder 4 (bottom). Columns: PSNR (dB, on the brightness channel), LPIPS and DISTS (both compare with the ground-truth image) and NIQE (needs no ground truth). ↑\uparrow / ↓\downarrow : higher/lower is better. Best per column over all methods in bold , second best underlined . The full nine-metric tables are in the full report.
More in this topic

Image Enhancement

Image enhancement restores visibility, color, detail, and dynamic range in degraded images for human viewing and downstream vision tasks.

See all 159 papers

Explore similar work

CardsList
Sep 21, 2026

ImIR: Image-Instruction Tuning for All-in-One Image Restoration

Degradations vary widely across images, so a practical restoration system has to handle many degradation types with one model. A recent and effective recipe adapts a large pretrained image-editing model to restoration using a small low-rank adapter with a text prompt. We replace that prompt with an instruction derived from the degraded image itself. The image reaches the editor through two paths: its structure comes from the model's VAE, and its semantic instruction comes from a lightweight token mapper that shifts the degraded image's vision-language embedding toward the embedding a clean image would produce. Because the instruction is a continuous vector, scaling it yields a family of valid restorations for tasks whose target is not unique, such as low-light enhancement. We adapt one Qwen-Image-Edit model to six tasks with a single adapter trained in about three hours on one GPU. The image instruction outperforms text conditioning under a matched comparison, and it supports task agnostic restoration without a degradation label, which the text variant does not.

Süleyman Aslan, Görkay Aydemir, Mısra Yavuz +5

Codeway AI Research

Visual Instruction TuningAll-in-One Image RestorationImage RestorationMultimodal Instruction TuningVLM Adaptation
AskPDF
Mar 3, 2026

MiM-DiT: MoE in MoE with Diffusion Transformers for All-in-One Image Restoration

All-in-one image restoration is challenging because different degradation types, such as haze, blur, noise, and low-light, impose diverse requirements on restoration strategies, making it difficult for a single model to handle them effectively. In this paper, we propose a unified image restoration framework that integrates a dual-level Mixture-of-Experts (MoE) architecture with a pretrained diffusion model. The framework operates at two levels: the Inter-MoE layer adaptively combines expert groups to handle major degradation types, while the Intra-MoE layer further selects specialized sub-experts to address fine-grained variations within each type. This design enables the model to achieve coarse-grained adaptation across diverse degradation categories while performing fine-grained modulation for specific intra-class variations, ensuring both high specialization in handling complex, real-world corruptions. Extensive experiments demonstrate that the proposed method performs favorably against the state-of-the-art approaches on multiple image restoration task.

Lingshun Kong, Jiawei Zhang, Zhengpeng Duan +6

Nanjing University of Science and Technology · Nankai University · Harbin Institute of Technology

All-in-One Image RestorationDiffusion Models for Image RestorationMixture of Experts
AskPDF
Mar 27, 2026

Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration

Image restoration aims to recover high quality images from inputs degraded by various factors, such as adverse weather, blur, or low light. While recent studies have shown remarkable progress across individual or unified restoration tasks, they still suffer from limited generalization and inefficiency when handling unknown or composite degradations. To address these limitations, we propose RAR, a Restore, Assess and Repeat process, that integrates Image Quality Assessment (IQA) and Image Restoration (IR) into a unified framework to iteratively and efficiently achieve high quality image restoration. Specifically, we introduce a restoration process that operates entirely in the latent domain to jointly perform degradation identification, image restoration, and quality verification. The resulting model is fully trainable end to end and allows for an all-in-one assess and restore approach that dynamically adapts the restoration process. Also, the tight integration of IQA and IR into a unified model minimizes the latency and information loss that typically arises from keeping the two modules disjoint, (e.g. during image and/or text decoding). Extensive experiments show that our approach consistent improvements under single, unknown and composite degradations, thereby establishing a new state-of-the-art.

I-Hsiang Chen, Isma Hadji, Enrique Sanchez +5

Samsung AI Center Cambridge · National Taiwan University · Technical University of Iasi +3

Image RestorationAll-in-One Image RestorationDegradation-Aware Image RestorationImage Corruption Robustness
AskPDF