When you snap a photo on a phone and it comes back with noise, blur, or washed-out details, you’re watching a quiet showdown between reality and its imperfect capture. Image restoration is the craft of mending that rift: turning a degraded image into something closer to what the scene actually looked like. For years, engineers chased better results by weaving local detail with global context—conjuring sharp edges from rough pixels, and guessing what belongs in unseen areas based on what’s nearby. Early efforts leaned on convolutional neural networks that excel at local neighborhoods but struggle to grasp long-range relationships. Then came vision transformers, which pair every pixel with every other, capturing global dependencies at a price: a lot of computation that grows quickly as images get bigger. The field has been stuck between speed and scope, until a new family of ideas borrowed from state-space modeling began to tilt the balance back toward both efficiency and understanding.
Headlining the latest advance is a collaboration between National Tsing Hua University in Taiwan, National Taiwan University, and MediaTek Inc., with Yu-Cheng Lin and Yu-Syuan Xu as the lead authors. They present Efficient All-Around Mamba, or EAMamba, an upgraded framework that keeps the ability to model long-range dependencies—crucial for restoring distant texture and structure—while dramatically cutting the computational cost. In other words, it’s the same quest as earlier attempts to “see the whole picture,” but it does so with a smarter map of where computation actually needs to be spent. The punchline is practical as well as theoretical: EAMamba can deliver strong restoration across a spectrum of tasks—super-resolution, denoising, deblurring, and dehazing—without the usual blowup in processing power. And the authors don’t just claim better numbers; they point to a pathway for real-world use in devices with limited resources.
What Vision Mamba tries to do
Vision Mamba, inspired by the Mamba state-space model, is designed to narrate the story of an image across long distances in a way that scales gracefully with the number of pixels. The core idea is intuitive: instead of examining every pixel in a vast grid in a brute-force, all-at-once fashion, you trace how information travels through a sequence, almost like following a thread through a tapestry. This allows the model to capture dependencies that stretch across the whole image, not just the nearby pixels that traditional local operators care about. And because it follows a state-space principle, the math can keep the computation linear in the image size, not quadratic. That’s a meaningful difference when you’re tearing through HD and 4K frames.
Yet even with that elegance, Vision Mamba faces practical friction. The way it historically aggregated different scanning directions—horizontal, vertical, and other patterns—meant the computation and the number of parameters could swell as researchers tried to capture more patterns. More importantly, there’s the issue of local pixel forgetting: in flattening a 2D image into a 1D stream, pixels that sit next to each other in the image can become distant in the sequence. That disconnect hurts the restoration task, which often relies on sharp, coherent local relationships at the heart of textures, edges, and subtle haze or blur cues. The field needed a design that could scale without multiplying the cost and, at the same time, preserve the neighborhood information that matters most for fidelity.
All-around scanning and MHSSM
Enter MHSSM, the Multi-Head Selective Scan Module. The trick is to split channels into groups and process each group through a tailored selective scan, then stitch the pieces back together. This channel-splitting trick keeps the computational burden in check while letting the model explore multiple sequences in parallel. Think of it as running several focused hunts for detail at once, rather than multiplying the length of a single chase. The outputs from each group are combined, preserving the richness of the signal without inflating the parameter count. The architecture also features a gating pathway that merges two streams—one that carries the scanned information and another that gates it through a nonlinear activation—via element-wise multiplication. The upshot is a more expressive, yet lean, token mixer inside the MambaFormer block.
But the real spark is the all-around scanning strategy. Where earlier approaches relied on two-dimensional scanning that could still miss information from certain directions, all-around scanning deliberately embraces multiple directions: horizontal, vertical, diagonal, flipped diagonal, and their reversed forms. The goal is to cover the neighborhood around every target pixel from many angles, reducing the chance that important context slips through the cracks. In visualizations of effective receptive field (ERF), the all-around approach shows stronger, more uniformly distributed influence from nearby and moderately far pixels, compared with simpler 2D scans. The authors argue that this holistic pattern is key to preserving local structure while still capturing global dependencies.
From a practical standpoint, MHSSM makes this all-around strategy affordable. By splitting channels, the method keeps the number of parameters and FLOPs roughly in line with simpler baselines, even as it aggregates information from multiple scanning directions. The authors quantify the payoff: EAMamba achieves a substantial reduction in FLOPs—ranging from 31% to 89% compared with prior Vision Mamba variants—without sacrificing performance. In other words, it’s a better engine for the same or even less fuel. It’s the difference between driving a car that sips gas versus one that sips less while delivering the same miles per gallon.
Why it matters beyond academia
Why should you care about a paper with many moving parts and a lot of technical detail? Because the implications ripple into devices we use every day. Image restoration isn’t just about prettier photos in your gallery; it’s about sharper videography on a phone, clearer frames in video calls, and more reliable vision in cameras and sensors across consumer electronics, robotics, and automotive systems. If you can restore high-quality visuals with far fewer floating-point operations, you can run more advanced models on phones and edge devices without draining the battery or overheating the chip. In environments where bandwidth or power is limited—think smartphones, drones, or autonomous machines—the ability to bring the same restoration quality with less computation can be transformative.
The paper’s results are broad: across denoising, real-world super-resolution, deblurring, and dehazing, EAMamba maintains competitive image quality while lowering computational load. In synthetic denoising benchmarks, for example, the authors report that EAMamba uses only about a tenth of the FLOPs of a strong Vision Mamba baseline yet remains very close in PSNR metrics. In real-world denoising on challenging datasets, their approach halved the FLOPs relative to comparable models while delivering comparable or slightly better visual fidelity. And in real-world super-resolution, EAMamba not only keeps the parameter count modest but also edges ahead on higher-RAISE inference tasks like ×4 scaling, where many models struggle to balance speed and accuracy. These aren’t memory-audit numbers on a whiteboard; they map to everyday experiences—faster photo edits, smoother video playback, and more robust performance in the field.
Beyond the numbers, the authors’ approach invites a shift in how we think about model design. The push toward channel-wise modularity and multi-directional scanning embodies a philosophy: you can gain global understanding by stitching together diverse, complementary views of the same data, without paying for a dogmatic, monolithic computation. It’s a little like how musicians layer motifs and rhythms from different instruments to create a richer piece without overloading the stage. This mindset—balancing expressive power with practical efficiency—could influence not just image restoration but other low-level vision tasks, and perhaps even real-time perception in robotics and vehicles.
Importantly, the authors make their code available, inviting the broader community to explore, test, and extend EAMamba. The repository can be found at the project page linked by the authors, reflecting a growing culture in which open-source collaboration accelerates the pace of improvement for tools that touch everyday life.
Where this could go next
If EAMamba represents a mature refinement of Vision Mamba, the next frontier is likely to be dynamic—video restoration, streaming pipelines, and real-time enhancement in consumer devices. Videos add a temporal dimension to the same scaling challenges: preserving coherence across frames while maintaining efficiency is an extra layer of complexity. The authors’ emphasis on all-around scanning could inspire temporal extensions, where the model learns to leverage not only spatial neighborhoods but short, computed temporal neighborhoods as well, all while keeping the linear-scaling promise intact.
There are still open questions. How well does the all-around strategy generalize to unseen degradations or novel camera pipelines? How low can FLOPs go before the fidelity regresses in ways that users would notice in everyday photos? And how can hardware—mobile GPUs, dedicated neural accelerators, or edge AI chips—best exploit the MHSSM’s modularity to deliver even faster restoration without sacrificing energy efficiency? These are not show-stopper questions but directions that researchers and industry partners will likely pursue in the coming years.
At the end of the day, EAMamba isn’t a single trick; it’s an architectural stance: controlled, scalable, and holistic. It shows that with careful design, a vision model can understand an image as a structured whole and still be light enough to run where power and space are at a premium. The work, originating from Taiwan’s leading universities and a major industry player, signals a maturation in the field—where researchers are not just chasing the highest PSNR curves but building practical systems that can live on devices, in the field, and in cameras we use every day.
