Andrii Bidochko

• Updated: July 21, 2026
• 6 min read

MambaLIE: Scene Light Intensity-Boosted Low-Light Image Enhancement with State Space Model

Direct Answer

MambaLIE introduces a novel low‑light image enhancement (LIE) pipeline that leverages a scene‑level light intensity cue together with a lightweight State Space Model (SSM) called the Locally Enhanced State Space Model (LESSM). By marrying long‑range illumination modeling with efficient local feature refinement, the method delivers state‑of‑the‑art visual quality while keeping compute and memory footprints low enough for deployment on smartphones and other edge devices.

Background: Why This Problem Is Hard

Consumer‑grade cameras—whether embedded in smartphones, wearables, or inexpensive digital cameras—are fundamentally limited by sensor size, pixel pitch, and on‑chip processing pipelines. In dim environments these constraints manifest as severe noise, color distortion, and loss of detail, which not only degrades user experience but also harms downstream computer‑vision tasks such as object detection, facial recognition, and autonomous navigation.

Two families of deep‑learning approaches dominate the LIE landscape today:

• Convolutional Neural Networks (CNNs) excel at extracting hierarchical patterns but operate on fixed‑size receptive fields. Consequently, they struggle to capture global illumination patterns that span the entire scene, leading to uneven brightness or halo artifacts.
• Vision Transformers (ViTs) address the receptive‑field limitation by attending to all patches, yet their quadratic attention cost balloons with image resolution. For high‑resolution mobile photography this translates into prohibitive latency and battery drain.

Practitioners therefore face a trade‑off: choose a fast but locally myopic CNN, or accept a globally aware transformer that cannot run in real time on resource‑constrained hardware. Bridging this gap—delivering both global illumination awareness and edge‑friendly efficiency—remains an open challenge.

What the Researchers Propose

The authors of MambaLIE paper propose a three‑pronged framework:

1. Scene Light Intensity Boosting: A pre‑processing step that estimates a coarse, scene‑wide illumination map. This map reshapes the distribution of light across the image, providing a global guide for subsequent enhancement.
2. State Space Model (SSM) Branch: An SSM processes the boosted illumination map to capture long‑range dependencies with linear time complexity, sidestepping the quadratic blow‑up of transformer attention.
3. Local Enhanced Branch: Parallel convolutional pathways refine fine‑grained textures, edges, and noise patterns, ensuring that the final output retains sharp local detail.

The two branches are fused through a gating mechanism that lets the scene‑level intensity signal dictate how much global versus local information each pixel should receive. The resulting architecture—named the Locally Enhanced State Space Model (LESSM)—delivers a balanced blend of global illumination consistency and local texture fidelity.

How It Works in Practice

At inference time, MambaLIE follows a deterministic pipeline:

1. Input Normalization: The raw low‑light image is first linearized and normalized to a common dynamic range.
2. Scene Light Intensity Estimation: A lightweight encoder predicts a single‑channel intensity map that reflects the overall lighting conditions of the scene. This map is then element‑wise multiplied (gated) with the normalized input, effectively “brightening” the image in a globally coherent manner.
3. Dual‑Branch Processing:
   • SSM Branch: The gated image passes through a recurrent‑style state space block. Because SSMs propagate information sequentially along a hidden state, they can model interactions across the entire image with O(N) complexity, where N is the number of pixels.
   • Local Enhanced Branch: Simultaneously, a shallow CNN extracts high‑frequency cues (edges, textures) and performs denoising using depth‑wise separable convolutions to keep the parameter count low.
4. Fusion & Output Reconstruction: A learned gating layer blends the SSM output and the local branch output, producing a final enhanced image that respects both global illumination balance and local detail preservation.

What sets this workflow apart is the explicit separation of “global lighting” and “local texture” concerns, each handled by a component that is mathematically optimal for its task. The SSM provides a linear‑time global context, while the convolutional branch supplies the non‑linear, high‑resolution refinements that CNNs excel at.

Evaluation & Results

The authors benchmarked MambaLIE on a suite of four synthetic low‑light datasets (e.g., LOL, SID) and five real‑world collections captured with consumer devices. Evaluation metrics included Peak Signal‑to‑Noise Ratio (PSNR), Structural Similarity Index (SSIM), and perceptual scores from a user study.

• Accuracy: Across all benchmarks, MambaLIE consistently outperformed leading CNN‑based methods (e.g., KinD++, RetinexNet) and matched or exceeded transformer‑based approaches (e.g., LLFormer) on PSNR/SSIM, often by margins of 0.5‑1.2 dB.
• Speed: On a mid‑range mobile SoC (Qualcomm Snapdragon 8‑Gen 2), the model processed a 1080p frame in under 45 ms, roughly 2‑3× faster than comparable transformers and on par with the fastest CNNs.
• Model Size: LESSM contains fewer than 4 M parameters, a fraction of the 30‑M‑plus parameter counts typical of ViT‑based LIE solutions, enabling on‑device deployment without external memory.

Qualitative visual comparisons highlighted smoother illumination gradients, fewer color shifts, and better preservation of fine details in night‑time street scenes and indoor low‑light portraits. The user study reported a 23 % higher preference rate for MambaLIE outputs over the next best baseline.

Why This Matters for AI Systems and Agents

Low‑light enhancement is more than a cosmetic upgrade; it directly influences the reliability of downstream AI agents that rely on visual input. Consider autonomous drones navigating at dusk, retail analytics cameras monitoring dimly lit aisles, or conversational agents that need to interpret user‑generated photos in chat. A robust LIE front‑end like MambaLIE can:

• Improve Detection Accuracy: Cleaner, well‑exposed frames boost object‑detection confidence scores, reducing false negatives in safety‑critical scenarios.
• Reduce Edge‑Compute Load: By delivering high‑quality enhancement with a sub‑4 M parameter model, MambaLIE frees CPU/GPU cycles for other inference tasks, enabling richer multimodal pipelines on the same hardware.
• Enable Real‑Time Interaction: Sub‑50 ms latency aligns with the responsiveness expectations of conversational agents and AR/VR experiences.

Developers building AI‑driven workflows can integrate MambaLIE into existing pipelines using the Workflow automation studio, allowing visual preprocessing to be orchestrated alongside language models, knowledge bases, and voice synthesis components.

What Comes Next

While MambaLIE marks a significant step forward, several avenues remain open for exploration:

• Dynamic Scene Adaptation: Current intensity boosting assumes a static illumination map. Future work could incorporate temporal consistency for video streams, reducing flicker in night‑time surveillance footage.
• Cross‑Domain Generalization: Training on synthetic benchmarks yields strong results, yet real‑world lighting conditions vary wildly. Semi‑supervised or self‑supervised fine‑tuning on device‑collected data could further close the domain gap.
• Hardware‑Specific Optimizations: Tailoring the SSM kernels to emerging AI accelerators (e.g., NPUs, Tensor Cores) could push latency below 20 ms for 4K video, opening doors for high‑resolution streaming applications.
• Integration with AI Agents: Embedding MambaLIE as a pre‑processing step in an Enterprise AI platform by UBOS would let large‑scale vision‑language agents automatically request illumination enhancement when confidence drops, creating a self‑healing perception loop.

Beyond research, the lightweight nature of LESSM makes it a prime candidate for inclusion in consumer‑facing products such as the ChatGPT and Telegram integration, where users could upload dark photos and receive instantly brightened results powered by on‑device inference.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.