🤖 AI

EvTexture++: Event-Driven Texture Enhancement for Video Super-Resolution

This paper introduces EvTexture++, the first event-driven framework that shifts the focus of event signals from motion refinement to texture enhancement in video super-resolution, utilizing a customized iterative enhancement branch and a temporal alignment module to achieve state-of-the-art performance and plug-and-play compatibility with existing models.

Original authors: Dachun Kai, Jiayao Lu, Yueyi Zhang, Xiaoyan Sun

Published 2026-06-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dachun Kai, Jiayao Lu, Yueyi Zhang, Xiaoyan Sun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Fixing Blurry Videos with "Super-Speed" Eyes

Imagine you are trying to watch a video of a fast-moving car, but the video is low-quality, blurry, and the details (like the text on the license plate or the texture of the car's paint) are completely lost. This is the problem Video Super-Resolution (VSR) tries to solve: turning a fuzzy, low-resolution video into a sharp, high-definition one.

Most current methods try to fix this by looking at the blurry pictures (frames) next to each other and guessing what the missing details should look like. However, if the car moves too fast, these methods get confused. They often end up "smearing" the image or making it look too smooth, like a painting rather than a photo.

EvTexture++ is a new solution that adds a special pair of "super-speed eyes" (called an Event Camera) to the process. These eyes don't take normal photos; instead, they only record changes in light at a speed so fast (microseconds) that they capture motion and texture details that normal cameras miss entirely.

The Core Problem: Texture vs. Motion

The authors noticed that previous methods using these "super-speed eyes" focused mostly on tracking motion (where the car is going). But they weren't very good at restoring texture (what the car actually looks like).

Think of it like this:

Old Methods: A detective who is great at tracking a suspect's footsteps (motion) but terrible at describing what the suspect's face looks like (texture).
EvTexture++: A detective who is great at both tracking the footsteps and reconstructing the face from the tiniest clues left behind.

How EvTexture++ Works: The Two-Tool Kit

The system uses two main tools working together, like a construction crew fixing a damaged wall:

1. The Texture Crew (The "Detail Hunter")

This part of the system is dedicated to bringing back the fine details (the "texture").

The Challenge: Event cameras are like a stream of individual sparks. They tell you something changed, but they don't give you the full picture of the wall's color or brightness.
The Solution: The system uses a clever trick called Iterative Texture Enhancement (ITE). Imagine trying to clean a dirty window. Instead of wiping it once, you wipe it, check the result, wipe it again, check again, and repeat.
- The system breaks the "spark" data (events) into tiny time-slices.
- It goes through these slices one by one, slowly adding more and more detail to the image.
- With every pass, the image gets sharper, recovering things like the stripes on a shirt or the leaves on a tree that were previously just a blur.

2. The Alignment Crew (The "Stabilizer")

When things move fast, the video can start to "flicker" or jitter.

The Challenge: If you try to stitch two blurry photos together while the camera is shaking, the result looks wobbly.
The Solution: This crew uses the "super-speed eyes" to track movement with extreme precision. Because event cameras react instantly to motion, they can see the movement of a fast object much better than a normal camera.
- The system uses this super-precise motion data to line up the frames perfectly before stitching them together.
- This stops the "flickering" effect, making the video look smooth and stable, even when the action is chaotic.

The "Plug-and-Play" Superpower

One of the coolest features of EvTexture++ is that it doesn't require you to rebuild the entire video player.

The Analogy: Imagine you have a high-end car engine (a modern video AI model). It's fast and powerful, but maybe the paint job is dull. EvTexture++ is like a turbocharger kit you can snap onto the engine.
You don't need to replace the engine. You just attach this new module, and suddenly, the car runs better and looks sharper.
The paper shows that they can take existing, state-of-the-art video models and "plug in" EvTexture++ to instantly make them much better at restoring details, without needing to retrain the whole system from scratch.

The Results: What Did They Find?

The researchers tested this on five different datasets (collections of videos).

Better than the Best: EvTexture++ beat all other current methods, including those that don't use event cameras and those that do.
Texture Richness: It performed especially well on videos with lots of details (like leaves, fabric patterns, or license plates). On one dataset called "Vid4," it improved the video quality by about 1.55 dB (a significant jump in video quality terms) compared to the previous best model.
Real-World Proof: They even tested it on real-world data (not just computer simulations) and it still worked better than the competition, proving it can handle the "noise" of real sensors.

Summary

EvTexture++ is a new tool that uses ultra-fast "motion sensors" (event cameras) to fix blurry videos. Instead of just guessing where things moved, it uses those sensors to rebuild the missing details (texture) and stabilize the movement simultaneously. It acts like a universal upgrade that can be attached to existing video AI models to make them see the world with much sharper, clearer eyes.

Technical Summary of EvTexture++: Event-Driven Texture Enhancement for Video Super-Resolution

Problem Statement
Video Super-Resolution (VSR) aims to reconstruct high-resolution (HR) video frames from low-resolution (LR) inputs by leveraging temporal dependencies. While recent advancements have integrated event camera signals to improve motion estimation and temporal alignment, existing event-based VSR methods primarily focus on motion refinement. Consequently, they often struggle with the restoration of fine-grained, high-frequency textures, leading to over-smoothed outputs and texture flickering in regions with complex details. Furthermore, standard RGB-based methods suffer from insufficient motion information due to low temporal sampling rates, hindering accurate alignment in large-motion scenarios. The core challenge addressed by this work is the specific restoration of texture details in VSR while maintaining temporal consistency, a task that remains under-explored despite the availability of event cameras which offer ultra-high temporal resolution and rich high-frequency dynamic details.

Methodology
The authors propose EvTexture++, the first event-driven framework dedicated explicitly to texture enhancement in VSR. The architecture is built upon a bidirectional recurrent structure (based on BasicVSR) and features two parallel, event-guided branches:

Event-Guided Texture Branch: This branch is designed to recover fine textural details by leveraging the high-frequency spatiotemporal information inherent in event signals.
- Iterative Texture Enhancement (ITE) Module: To address the challenge that events encode relative brightness changes rather than absolute intensity, the ITE module employs a voxel-based encoding of event streams. It utilizes a shared ConvGRU-based texture updater that iteratively refines features across $N$ iterations (where $N$ equals the number of voxel bins). By progressively injecting texture cues from voxel bins into the propagated features, the module effectively transfers high-frequency details from events to the RGB feature space, compensating for the lack of absolute intensity context in raw events.
Event-Guided Motion Branch: This branch focuses on enhancing temporal consistency, particularly in texture-rich regions where large motions often cause flickering.
- Temporal Texture Alignment (TTA) Module: The TTA module estimates optical flows from two complementary sources: an RGB-based flow (using a standard network like SpyNet) and an event-based flow (estimated via a U-Net processing voxelized events). The event-based flow is specifically "texture-aware," capable of capturing fast, non-linear motions with high temporal precision even in blurred regions. These two flows are fused to align cross-frame texture regions more accurately than using RGB flow alone.
Plug-and-Play Design: EvTexture++ is engineered as a flexible module that can be integrated into existing VSR backbones (ranging from CNNs to Transformers). In this configuration, the pretrained backbone remains frozen, and only the EvTexture++ modules and the upsampler are trained. This design allows for performance enhancement with minimal computational overhead and without requiring retraining of the entire backbone.

Key Contributions

First Dedicated Framework: The paper introduces the first event-driven framework specifically targeting texture restoration in VSR, shifting the focus from motion refinement to texture enhancement.
Progressive Recovery Mechanism: The proposal of a dedicated texture branch coupled with the ITE module, which iteratively refines spatiotemporal alignment and exploits sparse event hints to recover high-frequency details.
Temporal Consistency: The introduction of the TTA module, which estimates event-guided, texture-aware flow to mitigate texture flickering caused by large motions.
Modular Integration: The design of EvTexture++ as a universal plug-and-play tool that seamlessly integrates into diverse architectures (CNNs and Transformers), significantly boosting performance without altering the backbone.
Comprehensive Evaluation: Extensive experiments across five datasets (Vid4, REDS4, Vimeo-90K-T, UDM10, and CED) demonstrating state-of-the-art (SOTA) performance, particularly in texture-rich scenarios.

Results
Experiments demonstrate that EvTexture++ achieves SOTA performance across multiple benchmarks:

Quantitative Gains: On the texture-rich Vid4 dataset, EvTexture++ achieves a PSNR gain of up to 1.55 dB when integrated into the SOTA RGB-based model IART. Compared to the previous event-based SOTA (EBVSR), EvTexture++ improves PSNR by 0.27 dB on Vid4 and 0.16 dB on REDS4.
Perceptual Quality: The method achieves the lowest LPIPS score (0.2048) on Vid4, indicating superior perceptual quality and sharper edges compared to competing methods.
Robustness: The framework shows significant robustness under large motion (REDS4) and blur-downsampling (BD) degradation, outperforming both RGB-based and previous event-based methods.
Real-World Performance: On the real-world CED dataset, EvTexture++ achieves 33.71 dB PSNR, outperforming the best event-based alternative by 0.29 dB at 4× upsampling.
Efficiency: Despite the added complexity, EvTexture++ maintains a favorable trade-off between performance and efficiency, with a runtime of 95.0 ms per frame on 180×320 inputs, significantly faster than many Transformer-based competitors.

Significance
The paper positions EvTexture++ as a significant step forward in multi-modal VSR by explicitly addressing the under-explored challenge of texture restoration. By leveraging the unique properties of event cameras—specifically their high temporal resolution and high-frequency dynamic details—the framework overcomes the limitations of RGB-only methods in recovering fine structural details. The authors emphasize that their approach not only improves spatial fidelity but also enhances temporal stability, effectively reducing texture flickering. Furthermore, the plug-and-play nature of the framework offers a practical solution for upgrading existing VSR models, making high-quality texture recovery accessible without the prohibitive training costs associated with retraining large backbone architectures. The work validates that event signals, when properly processed through iterative refinement and temporal alignment, are highly effective for restoring complex textures in video sequences.