Revisiting Coarse-to-fine Paradigm in Nighttime Flare Removal via Visual Prompt

Dafeng Zhang

Keywords: Nighttime Flare Removal

Abstract: Flare removal is a crucial task in image processing, aiming to eliminate unwanted lens flare. Existing end-to-end flare removal methods, despite their progress, often introduce artifacts in the restored images. While multi-stage (coarse-to-fine) strategies in image restoration have proven effective for artifact suppression, their direct application to flare removal tasks yields limited improvements, raising questions about their inherent suitability. Inspired by inpainting techniques and prompt-based learning, we propose a plug-and-play Prompt Inpainting Network (PIN) that redefines coarse-to-fine processing for flare removal. We define this process as a Prompt Inpainting Pipeline (PIP). Our PIP introduces two synergistic mechanisms: Firstly, it leverages predicted flare mask from the coarse flare removal stage to explicitly exclude corrupted pixels and guide context-aware restoration. Second, high-quality decoder features from the coarse stage are repurposed as visual prompts to condition the refinement network, enabling feature-aware structural consistency in refinement stage. PIP is designed as a model-agnostic pipeline that seamlessly integrates with arbitrary restoration architectures, while introducing negligible computational overhead (minimum 1% parameters increment). Experiments demonstrate that PIP significantly reduces artifacts and achieves state-of-the-art performance across multiple benchmarks, proving that coarse-to-fine paradigms-when augmented with explicit corruption exclusion and visual prompts-are indeed effective for flare removal.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 5560