Simultaneously Achieving Performance and Efficiency in Low-light RAW Enhancement Architectures

Design and validate a deep learning architecture for low-light RAW image enhancement that simultaneously achieves strong restoration performance and high computational efficiency, with particular emphasis on suitability for resource-constrained devices by minimizing parameter count and FLOPs without sacrificing image quality.

Background

Low-light RAW image enhancement targets denoising and reconstruction directly in the RAW domain, which often involves high-resolution inputs and severe noise. While prior works such as Restormer, DNF, and RetinexRawMamba have advanced performance or efficiency individually, balancing both remains difficult—especially for single-image inference at full resolution.

This challenge is critical for deployment on mobile and edge devices where computation and memory budgets are limited. The paper motivates a rethinking of architectural design to harness complementary modules (Transformer-style channel attention for large-scale features and Mamba-based state-space modeling for small-scale features) to address this gap, underscoring that the general problem of achieving both strong quality and efficiency remains unresolved.

References

In particular, designing an architecture that achieves both strong performance and high efficiency is still an open problem. This issue is especially critical for resource-constrained devices, where lightweight models with fewer parameters and reduced FLOPs are essential.

— Rethinking Efficient Hierarchical Mixing Architecture for Low-light RAW Image Enhancement  (2510.15497 - Chen et al., 17 Oct 2025) in Section 1, Introduction (page 1)

The event visualization helps identify regions containing temporal contrast information, but its presence does not guarantee recovery when the RGB measurement is severely clipped or when the event signal is sparse or noisy. These examples motivate evaluating both pixel-level fidelity and perceptual quality and show that the current task remains unsolved despite the improvement over the baseline.

— SEE Challenge 2026: Event-Guided Brightness Adjustment Across a Broad Illumination Range  (2609.29347 - LU et al., 24 Sep 2026) in Section 3, Qualitative Observations; see also Figure 3 and the subsection “Shared Failure Pattern”

We conjecture that this issue could be alleviated if the model were able to understand the semantics of the scene. With such high-level understanding of what the image depicts and how a real-world scene should be structured, the model could infer the presence of these structures from weak signals, rather than ignoring them entirely. We leave the exploration of such a semantic-consistency-aware enhancement as future work.

— P-PatchDiff: Progressive Patch Diffusion Models for Low-light Image Enhancement  (2609.01123 - Guo et al., 1 Sep 2026) in Section 5.2, “Failure case analysis”