Research map / Vision-language and multimodal / Multimodal LLMs

Language-Guided One-Step Diffusion Model for Nighttime Flare Removal

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs

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 LG-ODM, a language-guided one-step diffusion framework for nighttime flare removal that uses a flare-specific VLM (Flare-VLM) to produce structured flare descriptions, a semantics-aware distribution distillation (SADD) loss, and an instruction-driven flare synthesis pipeline.

Problem

Nighttime flare removal from images with intense flare that fully occludes scene content; existing low-level restoration methods lack semantic priors to reconstruct occluded regions, causing artifacts and detail loss.

Approach

Fine-tune a flare-specific VLM to output a fixed-schema flare description (light source type, position, flare shape, color, occluded structure), convert it into a clean scene description that conditions a one-step diffusion student, and gate VSD noise injection spatially using a CLIPSeg flare mask so supervision concentrates on flare regions.

Key results

• On the authors' synthetic test set, Ours-1s reaches PSNR 31.844, SSIM 0.936, LPIPS 0.0527, FID 21.636, outperforming all listed baselines including CycleRDM-10s (FID 27.774).
• On the Flare7k++ test set, Ours-1s achieves PSNR 28.061 and FID 34.206 with 0.126 s inference on RTX 3090.
• On YOLOv11 detection of enhanced images, the model trained on the proposed data reaches mAP50 0.774 versus 0.742 without enhancement.

Surprising findings

• The H-F refinement stage yields only minor gains, suggesting most improvement comes from the one-step diffusion stage.

Limitations

• Reflective flares (e.g., aperture ghosts) are outside the synthesis scope.
• Relies on a separate high-frequency refinement stage to compensate for encoder-decoder high-frequency loss.

Experimental setup

Datasets: Flare7K++ ExDark DarkZurich low-light detection/driving/enhancement/segmentation datasets (about 12 sources)

Benchmarks: Flare7K++ test own synthetic test set Flare7K++ Real-world DarkZurich ExDark

Metrics: PSNR SSIM LPIPS FID NIQE MUSIQ CLIPIQA precision recall mAP50

Baselines: ConvIR Uformer BracketFlare Kotp et al. FPro AST Difflare CycleRDM OSEDiff

Builds on: OSEDiff Variational Score Distillation Flare7K++ CLIPSeg LoRA

Closest papers by meaning

• Disentangle Nighttime Lens Flares: Self-supervised Generation-based Lens Flare Removal (AAAI 2025 · similarity 0.88)
• FlareX: A Physics-Informed Dataset for Lens Flare Removal via 2D Synthesis and 3D Rendering (NeurIPS 2025 · similarity 0.86)
• Diff-Shadow: Global-guided Diffusion Model for Shadow Removal (AAAI 2025 · similarity 0.84)
• VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook (AAAI 2024 · similarity 0.83)
• FLARE: Feed-forward Geometry, Appearance and Camera Estimation from Uncalibrated Sparse Views (CVPR 2025 · similarity 0.82)
• See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.82)