Cameras play an essential role in autonomous driving, but their efficiency drops significantly at night. LightDiff addresses this by using a multi-condition controlled diffusion model to improve low-light images without paired human data, instead utilizing dynamic degradation.

Highlights:

• Tailored low-light image enhancement for autonomous applications.
• Novel adaptation of input weights from multimodal data including depth, RGB, and text.
• Context-consistent scene illumination maintaining visual integrity.
• Perception-guided reinforcement learning to align enhanced images with detection models.
• Proven efficacy on the nuScenes dataset with robust detection improvements in night-time conditions.

LightDiff is a critical innovation for the safety of autonomous driving at night. It underscores the potential for AI-driven enhancement to mitigate environmental constraints on vision-centric systems, and is particularly promising for models requiring improved perception under diverse lighting conditions.