Illumination Refinement via Textual Cues: A Prompt-Driven Approach for Low-Light NeRF Enhancement

In the realm of 3D scene modeling and rendering, the emergence of Neural Radiance Fields (NeRF) represents a significant leap forward. However, NeRF’s rendering performance suffers significantly when rendering images under low-light conditions. Existing approaches are optimized by enhancing low-light input images and combining NeRF models, but still fail to address the issues of multiview consistency and image quality. To address these challenges, our research introduces a textual constraint-prompted enhancement method that facilitates low-light image brightening and new view synthesis in an unsupervised manner. Specifically, we devise a semantic calibration strategy that employs positive and negative prompts to motivate and penalize the network towards attributes associated with high-quality images and exploits the capability of visual language models in semantic parsing to align the generated images with textual descriptors to improve image generation quality. In addition, to address the multiview consistency problem, we propose a two-layer optimization strategy, where the semantic cue optimization in the upper layer and the new view generation in the lower layer interact with each other to achieve a balance between luminance consistency and structural integrity by combining these improved images with text-driven semantic features. Comprehensive tests on two datasets with different resolutions, LOM and LLFF, show that our approach outperforms existing methods by significantly improving the brightness and clarity of low-light images to state-of-the-art while preserving the natural appearance and details.