Research map / Image generation / Image restoration
RIFLE: Removal of Image Flicker-Banding via Latent Diffusion Enhancement
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 RIFLE, a one-step latent diffusion (PiSA-SR-based) model with a flicker-banding prior estimator and masked loss, trained on a luminance-domain FB simulation pipeline and evaluated on a new paired real-world FB dataset.
Problem
Removing flicker-banding (alternating bright-dark stripes from rolling-shutter camera capture of emissive displays), an underexplored image degradation with no existing paired datasets.
Approach
Formulates FB removal as a dedicated restoration task: a simulation pipeline with jittered banding angle, spacing, width, feathered edges and sensor noise; a pretrained banding-parameter estimator injected into a one-step diffusion SR model; and a mask-weighted loss focusing supervision on banded regions.
Key results
- On the real-world Real-Flicker test set, RIFLE reaches PSNR 20.66 vs 20.57 for PiSA-SR and 19.43 for LQ input, with DISTS 0.1723 vs 0.1732.
- Ablation on the real-world set: ML+FPE reaches PSNR 22.30 vs 22.12 for the baseline, with GMSD 0.1286 vs 0.1297.
- Gains over PiSA-SR are small on most reference metrics; PiSA-SR has the best LPIPS (0.2389 vs 0.2456 for RIFLE).
Surprising findings
- Raw low-quality inputs already obtain relatively high reference-based scores, so metrics are largely insensitive to flicker-banding.
- Non-reference metrics were excluded because they gave similar scores for banded and banding-free images.
Limitations
- Gains over the PiSA-SR baseline on the real-world set are marginal on several metrics.
- Evaluation is limited to 105 image pairs from five scenes, cropped to 424 patches.
- Real-world test set release is not explicitly stated in the paper text (abstract mentions dataset release).
Experimental setup
Datasets: LSDIR UHDM
Benchmarks: Real-Flicker
Metrics: PSNR SSIM ms-SSIM LPIPS DISTS FSIM GMSD
Baselines: MAT InvSR PiSA-SR Step1X
Builds on: PiSA-SR LPIPS SAM2
Closest papers by meaning
- Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding (NeurIPS 2026 · similarity 0.93)
- BurstDeflicker: A Benchmark Dataset for Flicker Removal in Dynamic Scenes (NeurIPS 2025 · similarity 0.86)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.84)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.83)
- Disentangling Dual Image References in Frequency Aware Diffusion Models for Personalized Generation (NeurIPS 2026 · similarity 0.83)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.83)
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.