title	Flare Removal 2.0
emoji	🌖
colorFrom	purple
colorTo	purple
sdk	gradio
sdk_version	5.42.0
app_file	app.py
pinned	false
license	wtfpl

Demo Link 👉 - Flare Removal 2.0 – Hugging Face Demo

🌌 Lens Flare Removal – Research Implementations

This repository provides two end-to-end implementations of state-of-the-art lens flare removal methods:

• Wu et al., 2020 – How to Train Neural Networks for Flare Removal
• Zhou et al., 2023 (ICCV) – Improving Lens Flare Removal with General-Purpose Pipeline and Multiple Light Sources Recovery

Both tackle the same problem — removing lens flare artifacts to restore clean, high-quality images — but differ in their training data synthesis strategies and light source recovery mechanisms.

---

📖 Overview

Lens flare occurs when strong light sources (e.g., sun, street lights, car headlights) scatter or reflect inside a camera lens, producing streaks, blobs, or haze that degrade both visual quality and downstream computer vision tasks.

This project explores two influential research directions:

1. Wu et al. (2020) – Physics-inspired flare modeling (scattering + reflective) with additive composition and thresholded recovery.
2. Zhou et al. (2023) – ISP/AE-aware synthesis pipeline with smooth, threshold-free multi-source recovery.

---

🏗 Implementations

🔹 Wu et al. (2020) – Baseline

Key Ideas

• Flare Modeling:
  • Scattering flare – simulated via Fourier optics with random aperture defects.
  • Reflective flare – captured on a rotation stage with HDR imaging.
• Training Pair Synthesis:
  
• Losses: Combined image loss (L1 + perceptual) and residual flare loss.
• Light Source Handling: Mask saturated pixels (>0.99), ignore them in training, then feather original source back after inference.

Limitations

• Direct addition causes global brightening and clipping.
• Assumes one dominant light source.
• Limited cross-device generalization.

---

🔹 Zhou et al. (2023) – Improved Pipeline

Key Ideas

• ISP-Aware Data Synthesis:
  • Convex blending in inverse-gamma space (instead of direct addition).
  • Sigmoid weight map balances scene vs. flare, simulating auto-exposure darkening.
  • Noise sampled from .
• Convex Blending Equation:
  
• Threshold-Free Light Source Recovery:
  
  • Typically (\alpha = 15).
  • Recovers multiple emitters naturally, without threshold tuning.
• Evaluation: Tested on a Consumer Electronics dataset (10 devices, varied flare shapes).

Improvements Over Wu et al.

• More realistic synthesis → avoids distribution shift.
• Robust recovery → works for multiple light sources.
• Stronger cross-device generalization.
• Better downstream task performance (e.g., object detection after deflaring).

---

🔑 Key Differences

Aspect	Wu et al. (2020)	Zhou et al. (2023)
Flare Modeling	Physics-driven (sim + captured)	Focus on ISP/AE realism
Pair Synthesis	Direct addition in linear space	Convex blending in inverse-gamma space
Light Source Handling	Threshold mask + paste-back	Smooth power-law recovery (multi-source)
Generalization	Good for similar devices	Robust across devices & flare shapes
Focus	Flare physics	Camera pipeline + multi-source realism

---

⚙️ Usage

📦 Environment Setup

uv venv
.venv/Scripts/activate
uv add -r requirements.txt