Video Flare Dataset: Restoration & Forecasting

• Video Flare Dataset is a comprehensive video collection comprising synthetic and real-world sequences designed for advancing lens flare restoration and solar flare forecasting research.
• The dataset employs physics-informed synthesis and advanced sampling techniques, such as optical flow-based light-source motion, to generate realistic flare patterns across diverse scenes.
• Benchmarked with metrics like PSNR and SSIM, the dataset enables systematic evaluation of restoration algorithms and supports robust solar flare forecasting.

The Video Flare Dataset refers to both the synthetic and real-world video collections used to advance flare removal and flare forecasting methods in computer vision and solar physics. In recent literature, the term encompasses the VFlare dataset for lens flare restoration in natural scenes (Wang et al., 12 Dec 2025) and the HMI-video flare dataset for solar flare forecasting (Guastavino et al., 2021). Each constitutes a technically rigorous, class-balanced video corpus tailored to spatiotemporal flare dynamics under supervised and benchmarking protocols.

1. Dataset Composition and Scope

VFlare Dataset

The VFlare dataset comprises 1,500 synthetic paired video sequences derived from the REDS training set, each downsampled to 320×240 pixels. Scene content spans dynamic indoor and outdoor environments with camera and object motion, using strong punctual sources (e.g., headlights, sun, lamps) for flare simulation. Each sequence includes “clean” and “flare-degraded” counterparts; no further semantic or scene-type labels are provided. For testing and generalization, the dataset is augmented with dozens of unpaired real-world Internet videos that exhibit pronounced lens flare without ground-truth alignment. The test partition injects a static scattering-flare reference for enhanced light-source localization.

HMI-Video Flare Dataset

This dataset consists of videos of magnetograms from the Solar Dynamics Observatory’s Helioseismic and Magnetic Imager (SDO/HMI), spanning 2012–2017. Each clip samples 40 consecutive frames at 36-minute intervals (covering 24 h), resized to 128×128 pixels after projection to a cylindrical equal-area grid. Labels denote flare occurrence and class (X, M, C, NO1–NO4) based on GOES event catalogs, with detailed balancing to preserve archive class-proportions across training, validation, and test splits.

2. Synthetic Flare Generation and Labeling Protocols

VFlare Physics-Informed Synthesis

The synthetic flare pipeline leverages motion independence between flares, sources, and scene content:

• Light-source motion is governed by optical flow (FlowFormer++), with the position recursively updated by:

$$P_t^L(x, y) = P_{t-1}^L(x, y) + O_{t-1 \rightarrow t}(x, y)$$

• Scattering flare centers are offset from the light source:

$$P_t^s(x, y) = P_t^L(x, y) + t, \quad t \sim \mathcal{U}(0,15)$$

Shapes and orientations are sampled from aperture templates (circular, hexagonal, randomly rotated) and remain fixed temporally.

• Reflective flares are synthesized with collinearity constraints:

$$(x_3 - x_1, y_3 - y_1) = \lambda (x_2 - x_1, y_2 - y_1), \quad \lambda \in \mathbb{R}$$

Physically-based rendering modulates parameters such as element count and coating properties per scene, emulating phenomena described by Hullin et al.

Parameter ranges are broadly sampled: initial source positions span the full frame, scattering offsets in [0,15] px, aperture shapes randomly rotate through $[0,2\pi]$, and reflective/attenuative coefficients follow empirical distributions for mild to extreme flare severity.

HMI-Video Flare Construction and Labeling

Each AR clip is labeled by scanning the 24-hour post-clip interval for the maximum GOES flare class, with seven mutually exclusive class definitions (X/M/C/NO1–NO4) to facilitate both multi-class and Yes/No forecasting tasks. Selection protocols enforce AR-wise independence across splits, and sample proportions mirror five-year archive occurrence rates.

3. Real-World Data Acquisition and Augmentation

VFlare Real-World Supplement

Unpaired videos are curated from public sources like YouTube and Vimeo, prioritized for dynamic content with pronounced lens flares and moving light sources. These samples serve only qualitative generalization tests, lacking pixelwise ground-truth and requiring no alignment or manual annotation.

HMI Dataset: Augmentation Steps

Training-time augmentation comprises random horizontal/vertical flips and 90° frame rotations, with batch standardization applied after each convolution layer and $L_2$ regularization on kernels. Clips are resized to 128×128 pixels to facilitate model input uniformity.

4. Data Splits, Licensing, and Access

VFlare Splits and Release

The synthetic corpus is divided into 1,200 scenes for training (≈14,400 frames) and 300 for testing (≈3,600 frames). The validation set is not explicitly defined; a ~10% partitioning of training data is suggested. Dataset and code are available at https://github.com/BNU-ERC-ITEA/MIVF. Licensing terms are unspecified in the manuscript, necessitating consultation of the GitHub repository for usage details.

HMI-Video Splitting Paradigm

Each split (train: 3,000, val: 750, test: 750) is drawn AR-wise to prevent cross-split leakage; class-balanced sampling ensures representative test statistics. Clips and accompanying metadata are distributed as HDF5 containers (videos as $(40,128,128)$ arrays), with labels and flags for forecast thresholds. Source magnetogram archives reside at http://jsoc.stanford.edu, and curated video-label datasets plus code are on GitHub: https://github.com/mida-group/flare-video-dataset.

5. Benchmarking Protocols and Evaluation Metrics

VFlare Restoration Metrics and Baselines

Benchmarks rely on PSNR (full frame), PSNR-M (flare region masked), and SSIM metrics, compared across state-of-the-art video and image restoration architectures:

Method	PSNR (↑)	PSNR-M (↑)	SSIM (↑)
Input (degraded)	28.82	31.32	0.9436
BasicVSR++	40.59	36.71	0.9836
RVRT	40.22	36.51	0.9813
Restormer	41.10	36.00	0.9849
Flare7K++†	29.65	31.74	0.9491
DeflareMamba†	30.58	32.15	0.9533
MIVF (Ours)	41.54	36.71	0.9861

† Indicates model fine-tuned on VFlare. The MIVF approach demonstrates the highest PSNR and SSIM, with restoration outperforming frame alignment-based methods and suppressing flicker in qualitative real-world evaluations.

HMI-Video Flare Forecasting Metrics and Splitting Rationale

Dataset splits support robust estimation of mean and standard deviation for the True Skill Statistic (TSS) over held-out sets, with randomization performed 10 times for independent evaluation. The balancing paradigm enables fair, reproducible benchmarking of deep learning models for flare classification and forecasting.

6. Significance and Applications

The VFlare dataset is the first large-scale video corpus for lens flare restoration, designed with motion-independence principles to facilitate rigorous benchmarking of flare removal algorithms. Its synthetic–real structure permits assessment of generalization for both quantitative and qualitative restoration tasks. The HMI-video flare dataset sets a paradigm for supervised solar flare event forecasting using temporal magnetogram sequences, ensuring class-proportionality and AR independence for valid machine learning evaluations. These resources collectively enable systematic comparison of model architectures, loss functions, and data-driven approaches for spatiotemporal flare mitigation and event prediction, catalyzing advancement in vision restoration and heliophysics research.