Bricker: A Bracketed-Exposure RAW Dataset for Flicker-Banding Removal

Paper: Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding (NeurIPS 2026) · arXiv:2606.29845 Code: github.com/ZZH-qwq/BRACE

Flicker-banding (FB) is the striped, colour-shifted pattern that appears when a rolling-shutter camera photographs a screen whose brightness is modulated by PWM or progressive refresh. The artifacts are content-dependent, and their morphology changes drastically with exposure — so a single frame can be nearly impossible to separate from real texture.

Bricker is the first bracketed-exposure RAW dataset built for this problem. Each sample is a five-frame exposure ladder of the same screen content plus an almost artifact-free ground truth, stored in the RAW domain. It comes in two paired subsets that share an identical file layout, dtypes and metadata keys, so one loader works for both:

subset	how it is made	purpose
Bricker-Syn	ray-traced physical simulation of screen imaging, with randomized 6-DOF screen pose, display refresh schemes, rolling-shutter speed and sensor noise	large-scale pretraining
Bricker-Real	automated multi-exposure RAW capture of real screens (Xiaomi 17 Pro + two LED matrix displays)	real-world fine-tuning and evaluation

Dataset statistics

subset	Train	Test	stored RAW grid (Hp × Wp)	rendered RGB PNG
syn	1000	100	512 × 512	1024 × 1024
real	250	40	384 × 512 or 512 × 384	768 × 1024 or 1024 × 768

• 1,390 samples × (5 bracketed frames + 1 ground truth) = 8,340 frame pairs.
• 18,070 files, 30.6 GiB total (≈ 18.0 GiB RAW arrays, ≈ 12.6 GiB RGB PNG renders).
• Train and test use disjoint screen content. The real train/test sets share the same capture device and the same pair of displays, so real fine-tuning can be validated on the same domain.

Exposure ladder

Frame i of a sample has exposure time T_i = 4^(i-3) · T_ref, i.e. underexposed → reference → overexposed. I3 is the reference frame; the ground truth is captured with a long exposure and a proportionally lower ISO so that banding averages out while brightness is preserved.

file	exposure	role
raw_I1.npy	1/16×	most severe banding
raw_I2.npy	1/4×	strong banding
raw_I3.npy	1×	reference frame
raw_I4.npy	4×	mild banding
raw_I5.npy	16×	nearly banding-free, limited dynamic range
raw_gt.npy	8× + ISO/8	artifact-free ground truth

Directory layout

<dataset-root>/
├── data/
│   ├── real/{Train,Test}/scene_XXXX/<sample>/  # raw_I1..I5.npy, raw_gt.npy, metadata.npy
│   │   └── ...                                 # + rgb_I1..I5.png, rgb_gt.png (13 files/sample)
│   └── syn/{Train,Test}/scene_XXXX/<sample>/   # same layout, same 13 files per sample
├── scripts/                                    # loading / visualisation / Dataset helpers
├── tools/                                      # network-ISP exporter used for the RGB renders
└── requirements.txt

<sample> is NNNNNN_<content-type>, e.g. 000011_web1; the numeric prefix is unique inside its subset, and the suffix marks the screen content:

suffix	content	syn Train / Test	real Train / Test
static	natural scenes	350 / 35	31 / 5
game	video games	200 / 20	46 / 7
web1, web2	web pages	200 / 20	57 / 9
doc1, doc2	documents	150 / 15	58 / 9
UI	mobile UI	50 / 5	29 / 5
anime	cartoon	50 / 5	29 / 5

Content was drawn from existing screen-content corpora (LSDIR, DocBank, PubLayNet, danbooru2021, VideoGameBunny, Rico, and public webpage collections) and filtered by an IQA check; see the paper for details and citations.

Per-sample files

file	dtype	shape	notes
raw_I1.npy … raw_I5.npy	uint16	(Hp, Wp, 4)	values are round(normalized × 1023), each frame normalized to its own white level
raw_gt.npy	float32	(Hp, Wp, 4)	flicker-free reference, 1.0 = white level
metadata.npy	dict	—	per-sample RAW/camera parameters (see below)`
rgb_I1.png … rgb_I5.png, rgb_gt.png	uint8 RGB	(2Hp, 2Wp, 3)	8-bit display-referred renders

The four channels are a 2×2-packed Bayer mosaic in R, G1, G2, B order (RGGB, channel 0 at even/even positions). All five frames of a sample share the same (Hp, Wp); sizes differ between samples.

metadata.npy keys: LinearizationTable, BlackLevel, WhiteLevel, CFAPattern, Orientation, rgb2cam, cam2rgb, rgb_gain, red_gain, blue_gain, AsShotNeutral, ColorMatrix1, ColorMatrix2.

Getting started

pip install -r requirements.txt
huggingface-cli download ZZH-qwq/Bricker --repo-type dataset --local-dir Bricker

import numpy as np

sample = "Bricker/data/real/Test/scene_0000/000000_game"
frames = {k: np.load(f"{sample}/raw_I{k}.npy") for k in range(1, 6)}  # uint16 (Hp, Wp, 4)
gt = np.load(f"{sample}/raw_gt.npy")                                   # float32 (Hp, Wp, 4)
meta = np.load(f"{sample}/metadata.npy", allow_pickle=True).item()

x = frames[3].astype(np.float32) / 1023.0   # reference frame in [0, 1]
r, g1, g2, b = (x[..., i] for i in range(4))

Ready-made helpers are in scripts/:

python scripts/load_sample.py data/real/Test/scene_0000/000000_game   # inspect one sample
python scripts/visualize.py   data/real/Test/scene_0000/000000_game   # write preview_I3.png
python scripts/dataset.py --subset real --split Train                 # PyTorch Dataset smoke test

RGB renders

Every sample ships six 8-bit PNG renders produced by a network-based ISP followed by a display curve: a μ-tonemap (μ = 5000) for the real subset and 2.2 gamma for the synthetic subset. They are conveniences for inspection and for metrics computed in sRGB space; the RAW .npy files are the authoritative data.

To regenerate them, see tools/ — it contains the trimmed ISP pipeline and the 2.2 MB demosaic network weights:

python tools/export_rgb.py --subset real --limit 2 --device cpu   # check
python tools/export_rgb.py --subset all --device cuda:0            # regenerate everything

Intended use and limitations

• Intended: training and benchmarking multi-frame / bracketed RAW restoration for screen-capture artifacts — flicker-banding removal, cross-exposure fusion, and RAW-domain ISP research.
• The real subset is deliberately small (290 samples from 4 capture scenes) and uses one phone and two LED matrix displays; it validates real-domain fine-tuning, not device-general performance.
• Real captures use controlled lighting (no external reflections or glare) because an artifact-free ground truth requires long exposures.
• Synthetic samples come from a physically simulated pipeline and therefore still show a domain gap to real captures — this gap is exactly what the paired real subset is meant to measure.
• Screens, UI and document content originate from public corpora; please also respect the terms of those upstream sources.

Licensing

The dataset is released under CC BY-NC-SA 4.0: non-commercial use with attribution and share-alike — see LICENSE; scripts/ uses the same terms. tools/bayer_net/ (the demosaic network and its weights) is from demosaicnet under MIT.

Citation

@article{zhou2026bricker,
  title   = {Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding},
  author  = {Zhou, Zihan and Zhu, Libo and Gong, Jue and Zhou, Zhiyi and Cao, Jiezhang and Guo, Yong and Zhang, Yulun},
  journal = {Advances in Neural Information Processing Systems},
  year    = {2026}
}

Contact: Zihan Zhou (Shanghai Jiao Tong University) — issues and questions are welcome on the code repository.