PromptLNet: Region-Adaptive Aesthetic Enhancement via Prompt Guidance in Low-Light Enhancement Net

TL;DR

This work tackles the mismatch between objective metrics and aesthetic quality in low-light image enhancement by introducing a prompt-driven, region-adaptive framework guided by human aesthetic preferences via RLHF. It combines a Retinex-based reasoning segment, region-specific brightness controls, adaptive contextual compensation, and diffusion generation controlled by ControlNet, all steered by an aesthetic reward model trained with BLIP-based features. A PFNet-driven pipeline automates the aesthetic alignment process, supported by a curated dataset of 1000 samples with multi-dimensional expert and automated scores across diverse lighting conditions. Experimental results demonstrate improved visual quality and controllability over existing methods across multiple benchmarks, highlighting the potential for practical, user-driven aesthetic refinement in real-world applications. Overall, the approach advances low-light enhancement by aligning光image outputs with human preferences while enabling semantic, region-level control.

Abstract

Learning and improving large language models through human preference feedback has become a mainstream approach, but it has rarely been applied to the field of low-light image enhancement. Existing low-light enhancement evaluations typically rely on objective metrics (such as FID, PSNR, etc.), which often result in models that perform well objectively but lack aesthetic quality. Moreover, most low-light enhancement models are primarily designed for global brightening, lacking detailed refinement. Therefore, the generated images often require additional local adjustments, leading to research gaps in practical applications. To bridge this gap, we propose the following innovations: 1) We collect human aesthetic evaluation text pairs and aesthetic scores from multiple low-light image datasets (e.g., LOL, LOL2, LOM, DCIM, MEF, etc.) to train a low-light image aesthetic evaluation model, supplemented by an optimization algorithm designed to fine-tune the diffusion model. 2) We propose a prompt-driven brightness adjustment module capable of performing fine-grained brightness and aesthetic adjustments for specific instances or regions. 3) We evaluate our method alongside existing state-of-the-art algorithms on mainstream benchmarks. Experimental results show that our method not only outperforms traditional methods in terms of visual quality but also provides greater flexibility and controllability, paving the way for improved aesthetic quality.

Figures (11)

• Figure 1: Overview of tasks to improve low-light conditions. The figure below illustrates the three levels of image enhancement under low-light conditions: (a) global enhancement, which is a uniform brightness adjustment of the entire image;(b) Region-level enhancement, which selectively enhances certain parts of the image, such as the background or objects of interest; (c) Prompt-driven enhancement, which uses natural language prompts to adjust the target area or the entire image under the guidance of semantic understanding.
• Figure 2: The overview of our framework, including the RRS Module, BC Module, ACC Module, Color Module, and control diffusion.
• Figure 3: Examples of some images from the preprocessed DCIM and MEF datasets. These images exhibit significant differences in aspects such as hue, brightness, and saturation.
• Figure 4: Examples of images from the LIME dataset along with their corresponding human feedback scores.
• Figure 5: Step-by-step image enhancement process using iterative adjustments and reward-based evaluation in color module. The workflow begins with the original image (Reward Score: -2.82), where brightness is increased by 20%, leading to a minor improvement in the reward score (-1.34). Subsequent adjustments include further brightening by 50% and boosting saturation by 25%, progressively enhancing the visual appeal while improving the reward score to -0.93. However, reducing saturation by 10% results in a slight decline in the reward score (-1.12). Finally, a slight yellow tone adjustment achieves the optimal reward score of 0.22, indicating significant aesthetic and perceptual improvements throughout the iterative process. This demonstrates the effectiveness of the framework in balancing clarity, color quality, and aesthetic appeal.