XIVE: Explainable and interactive methods for image and video enhancement

XIVE: Explainable and interactive methods for image and video enhancement

Artificial intelligence has transformed the way we create, edit and consume images and videos. From social media filters to the restoration of old photographs and low-cost audiovisual production, visual editing tools are now more accessible and powerful than ever. The growing demand for intelligent and user-friendly editing tools requires solutions capable of addressing both traditional challenges, such as color balancing, noise reduction and low-light image enhancement, as well as more advanced techniques, including style transfer and automatic image and video editing.

However, this evolution also raises new challenges: how can we ensure that results are reliable? How can we understand the decisions made by a neural network? And how can we enable users to play a more active role throughout the creative process? These are some of the questions that have guided the XIVE project throughout its four years of research.

Led by the Computational Color research group at CVC and headed by Dr. Javier Vázquez-Corral, the project has investigated new techniques for image and video enhancement, restoration and editing. The aim has been to develop more explainable, interactive and transparent artificial intelligence systems, capable of giving users greater control over the results and a better understanding of how they are generated.

New tools for visual enhancement and editing

Throughout the project, the research team has developed a range of innovative solutions for image and video processing, with applications in areas as diverse as digital heritage restoration, computational photography, audiovisual content creation and professional editing tools.

One of XIVE’s main objectives has been to give users a more active role in the editing process. Rather than allowing artificial intelligence to make all decisions automatically, the project has explored new ways to enable people to better understand, guide, and control the behavior of algorithms.

Some of the most notable applications developed within the project include:

  • NamedCurves: Learning a photographer’s style to apply it to new images

Among XIVE’s most significant outcomes is NamedCurves, a technique capable of learning the editing style of a professional photographer and automatically reproducing it on new images.

The method works by separating an image into different colors that can be identified by the user (such as the blue of the sky or the green of vegetation) and learning how to adjust each of them independently. These adjustments are then combined to reproduce the desired visual style.

This approach enables image enhancement in an intuitive and understandable way, facilitating creative control over the result.

| More information: NamedCurves

| Paper: Serrano-Lozano, D., Herranz, L., Brown, M. S., & Vazquez-Corral, J. (2024). NamedCurves: Learned Image Enhancement via Color Naming. ECCV 2024.

  • HyperNVD: Faster video editing with AI

Another of the project’s advances has been HyperNVD, a technology designed to accelerate AI-based video editing. The tool automatically decomposes a video into different editable layers, making it possible to modify specific elements of a scene independently. This capability is particularly useful for creative applications and audiovisual production.

By leveraging hypernetworks and meta-learning techniques, HyperNVD significantly reduces the time required to process new videos, enabling results to be obtained much more efficiently than with traditional methods.

| More information: HyperNVD: Accelerating Neural Video Decomposition via Hypernetworks

| Paper: Pilligua, M., Xue, D., & Vazquez-Corral, J. (2025). HyperNVD: Accelerating neural video decomposition via hypernetworks. CVPR 2O25.

The advances developed within the XIVE project have led to numerous scientific publications in some of the leading international conferences and journals in computer vision and artificial intelligence, including ECCV, CVPR, AAAI, WACV, CIC and ICASSP.

The results cover areas such as low-light image enhancement, color correction, image restoration, object insertion, visual consistency across multiple images, scene editing and relighting, image compression, and visual quality metrics. In this way, XIVE demonstrates that it is possible to develop artificial intelligence systems that not only generate high-quality results, but are also more understandable, transparent and interactive.

Project PID2021-128178OB-I00 funded by MICIU/AEI/10.13039/501100011033 and by FEDER/EU.