Research map / Image generation / Image restoration
Light the Night: A Multi-Condition Diffusion Framework for Unpaired Low-Light Enhancement in Autonomous Driving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LightDiff, a ControlNet-style multi-condition Stable Diffusion model trained on synthetically degraded nuScenes daytime images, with LiDAR-depth and MMD distribution rewards, to enhance nighttime camera images for 3D car detection.
Problem
Vision-centric autonomous driving perception degrades in low-light conditions, and collecting paired nighttime-daytime driving data is costly and impractical.
Approach
Generate synthetic day-night pairs via an online low-light degradation pipeline (sRGB-to-RAW, attenuation, shot/read noise, ISP), and condition Stable Diffusion on a degraded image, depth map and caption through a multi-condition adapter that softly weights modalities, then fine-tune with LiDAR-depth and MMD distribution rewards.
Key results
- On nuScenes night validation, BEVDepth AP rose from 0.134 to 0.176 (+4.2%) with LightDiff-enhanced front-camera images.
- BEVStereo AP improved from 0.124 to 0.170 (+4.6%) with enhanced images.
- LightDiff achieved best scores on the four no-reference IQA metrics reported (e.g., MUSIQ 51.674 vs 40.695 for ShadowDiffusion).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enhance nighttime images with LightDiff | BEVDepth · nuScenes nighttime validation | AP: 0.134 -> 0.176 | Helped |
| enhance nighttime images with LightDiff | BEVStereo · nuScenes nighttime validation | AP: 0.124 -> 0.170 | Helped |
| add LiDAR and Distribution Reward Modeling (LDRM) (ablation) | LightDiff · nuScenes nighttime validation | AP: 0.151 -> 0.176 | Helped |
| add the Multi-Condition Adapter (ablation) | LightDiff · nuScenes nighttime validation | AP: 0.152 -> 0.176 | Helped |
| add Recurrent Lighting Inference (ReLI) (ablation) | LightDiff · nuScenes nighttime validation | AP: 0.163 -> 0.176 | Helped |
| add text and depth-map inputs to image-only input (ablation) | LightDiff · nuScenes nighttime validation | AP: 0.119 -> 0.176 | Helped |
| add text input (ablation) | LightDiff without text · nuScenes nighttime validation | AP: 0.146 -> 0.176 | Helped |
| add depth-map input (ablation) | LightDiff without depth map · nuScenes nighttime validation | AP: 0.139 -> 0.176 | Helped |
Surprising findings
- Several enhancement methods (SCI, Zero-DCE++, URetinex-Net) reduced 3D detection AP relative to raw night images, showing visual enhancement does not guarantee better perception.
Limitations
- Evaluation relies on no-reference IQA metrics because no paired day-night real data exists.
- Only the front camera is enhanced and evaluated, on a single dataset (nuScenes) with 15 nighttime validation scenes.
- Synthetic degradation may not capture real night lighting such as vehicle lights.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: MUSIQ HyperIQA MANIQA NIMA TReS ILNIQE NIQE AP ATE ASE
Baselines: Zero-DCE++ URetinex-Net SCI ShadowDiffusion ExposureDiffusion SNR-Aware EnlightenGAN CLIP-LIT BEVStereo
Builds on: Stable Diffusion ControlNet Zero-DCE++ BEVDepth
Closest papers by meaning
- UniDrive: Towards Universal Driving Perception Across Camera Configurations (ICLR 2025 · similarity 0.85)
- DriveGEN: Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation (CVPR 2025 · similarity 0.85)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.84)
- A New Multi-Source Light Detection Benchmark and Semi-Supervised Focal Light Detection (NeurIPS 2024 · similarity 0.84)
- Effective Multi-sensor Conditioning for Street-view Novel-view Synthesis (NeurIPS 2026 · similarity 0.84)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.