UniCoRN : Latent Diffusion-based Uni fied Co ntrollable Image R estoration N etwork across Multiple Degradations
Debabrata Mandal, Soumitri Chattopadhyay, Guansen Tong, Praneeth Chakravarthula · 2025
Read UniCoRN : Latent Diffusion-based Uni fied Co ntrollable Image R estoration N etwork across Multiple Degradations by Debabrata Mandal, Soumitri Chattopadhyay, Guansen Tong, Praneeth Chakravarthula (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-15868.txt.
Introduction
Debabrata Mandal, Soumitri Chattopadhyay, and colleagues build one restoration network that repairs many kinds of image damage in a 2025 computer vision study. Blur, noise, and haze usually each demand their own specialist model, yet real photos arrive with several faults stacked together. A unified system also needs controls, so a user can say which fault to fix and how hard.
Working in a latent diffusion framework, their network handles multiple degradations under one set of weights. Controllable, all-in-one restoration fits consumer photo tools and archives where the damage mix is never known in advance.
Catalog record
| CREATOR | Debabrata Mandal, Soumitri Chattopadhyay, Guansen Tong, Praneeth Chakravarthula |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-15868
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
UniCoRN : Latent Diffusion-based Uni fied Co ntrollable Image R estoration N etwork across Multiple Degradations Debabrata Mandal, Soumitri Chattopadhyay, Guansen Tong, Praneeth Chakravarthula Abstract Image restoration is essential for enhancing degraded images across computer vision tasks. However, most existing methods address only a single type of degradation (e.g., blur, noise, or haze) at a time, limiting their real-world applicability where multiple degradations often occur simultaneously. In this paper, we propose UniCoRN , a unified image restoration approach capable of handling multiple degradation types simultaneously using a multi-head diffusion model. Specifically, we uncover the potential of low-level visual cues extracted from images in guiding a controllable diffusion model for real-world image restoration and we design a multi-head control network adaptable via a mixture-of-experts strategy. We train our model without any prior assumption of specific degradations, through a smartly designed curriculum learning recipe. Additionally, we also introduce MetaRestore , a metalens imaging benchmark containing images with multiple degradations and artifacts. Extensive evaluations on several challenging datasets, including our benchmark, demonstrate that our method achieves significant performance gains and can robustly restore images with severe degradations. Project page: https://codejaeger.github.io/unicorn-gh Figure 1 : We present UniCoRN , a unified image restoration approach for handling multiple degradations simultaneously. Our approach shows robust performance on existing benchmarks and excels on MetaRestore , exhibiting multiple degradations. UniCoRN , based on stable diffusion, is corruption-agnostic, flexible, and scalable. 1 Introduction Image restoration [ 77 , 2 , 95 , 94 , 67 ] is a fundamental ill-posed inverse problem [ 2 , 94 , 67 ] in computer vision aimed at recovering high-quality images from its degraded, lower-quality counterparts. These degradations often arise during image acquisition due to undesirable environmental conditions [ 38 , 49 ] (e.g., low-light, haze, rain, blur) or inherent camera limitations [ 95 , 85 , 21 ] (e.g., sensor noise, chromatic aberrations), and can impact downstream applications. Most image restoration approaches have focused on addressing a single type of degradation, such as denoising [ 20 ] , dehazing [ 38 , 76 ] , deblurring [ 17 , 96 ] , deraining [ 94 , 44 ] , or low-light enhancement [ 49 , 62 ] . In real-world scenarios, imaging systems often capture images in unconstrained, dynamic environments where multiple degradations may occur simultaneously, such as autonomous vehicles facing multiple weather-related degradations [ 85 ] , underwater images affected by light scattering and absorption [ 99 , 35 ] , or images from low-quality cameras [ 85 , 23 , 28 ] or flat optics [ 36 , 80 , 79 , 9 ] prone to multiple corruptions . Unfortunately, the existing approaches [ 94 , 67 , 44 ] are mostly tailored for removing sole presence of individual degradation types and do not generalize well to handle multiple real-world corruptions simultaneously. We ran a simple experiment of applying degradation-specific models [ 94 , 67 ] sequentially to remove simultaneous presence of multiple artifacts. As shown in Fig. 2 , both approaches struggle to accurately restore the original image, often compounding artifacts and losing finer details. This highlights the limitations of single degradation removal methods, and necessitates simultaneous multi-degradation removal image restoration approaches. To address the gap in multi-degradation image restoration, we propose UniCoRN , a unified model that simultaneously handles various types of degradations through robust artifact removal and high-fidelity image preservation. Leveraging controllable diffusion [ 72 ] models as powerful generative priors [ 29 ] , we model image restoration as a conditional generation task. Inspired by ControlNet [ 97 ] , we use a multi-head control network where each head targets a specific degradation type commonly seen in real-world images [ 95 ] . The proposed UniCoRN architecture allows us to utilize low-level 1 1 1 Low level vision refers to using image features such as edges, contours, color maps, transmission maps and depth maps visual features [ 14 , 26 , 93 , 84 ] as degradation cues that condition image restoration and generation, providing better control than simple textual prompts [ 97 , 60 ] . To be precise, a non-exhaustive set of low-level visual “hints”, such as transmission, color or edge maps [ 26 , 31 , 62 ] , form the conditions for our control network, while a fixed generative diffusion model [ 72 ] serves as a backbone. Furthermore, a mixture-of-experts [ 74 , 8 ] strategy adaptively weights degradation cues, and curriculum learning [ 3 ] enables the model to progressively learn handling both single and compound degradations, especially mitigating the risk of catastrophic forgetting [ 15 , 34 ] across tasks. We have evaluated UniCoRN on diverse single and multi-degradation image restoration datasets [ 49 , 61 , 40 , 55 , 1 ] , and demonstrate significantly improved performance over prior state-of-the-art methods [ 94 , 30 , 67 , 43 , 101 ] . Additionally, we introduce MetaRestore , a new benchmark dataset with metalens-captured images [ 80 ] containing multiple real-world degradations and we further validate our model’s robustness to real-world in-the-wild unconstrained photography challenges. Figure 2 : Applying a sequential combination of single degradation image restoration models still struggle to faithfully recover an image corrupted by multiple degradations. UniCoRN , on the other hand, can recover from multiple degradations, without prior knowledge of the corruption type. Our core contributions, in summary, are as follows: 1. We introduce UniCoRN , a controllable latent diffusion model for unified image restoration that effectively recovers images with multiple, unknown degradations. Unlike most previous methods, our approach does not require prior knowledge of the specific degradations present. 2. We design a multi-head control network that utilizes diverse low-level visual cues, adapting these through a mixture-of-experts strategy. We train the network with a curriculum learning approach to prevent catastrophic forgetting across various degradation removal tasks. 3. We propose a new benchmark, MetaRestore , consisting of clean and metalens-captured degraded image pairs, to advance multi-degradation image restoration research. 4. We demonstrate that our approach not only outperforms prior methods but also generalizes robustly to our new metalens imaging benchmark, achieving high efficacy in addressing diverse and challenging artifacts. Code and data will be made public upon acceptance. 2 Related Works Image Restoration for Single Degradations: A bulk of image restoration approaches have focused on single degradation removal , i.e. when an image is assumed to have only one form of (known) degradation. These include denoising [ 7 , 39 , 20 , 31 ] , dehazing [ 76 , 18 , 42 , 38 , 41 , 88 ] , deblurring [ 96 , 17 , 78 , 90 , 31 ] , low-light enhancement [ 49 , 62 , 71 , 91 , 54 ] , among others. Traditional methods have relied on color [ 31 ] , dark channel [ 88 , 26 ] , bright channel [ 62 ] and low-rank [ 71 ] priors, which are all specific to a given degradation type and hence they cannot be applied across different degradations. More recently, deep learning models for image restoration [ 76 , 94 , 18 , 41 , 86 , 46 ] have emerged as superior alternatives by enabling