Capturing screens is now routine in our everyday lives. But the photographs of emissive displays are often influenced by the flicker-banding (FB), which is alternating bright%u2013dark stripes that arise from temporal aliasing between a camera's rolling-shutter readout and the display's brightness modulation. Unlike moire degradation, which has been extensively studied, the FB remains underexplored despite its frequent and severe impact on readability and perceived quality. We formulate FB removal as a dedicated restoration task and introduce Removal of Image Flicker-Banding via Latent Diffusion Enhancement, RIFLE, a diffusion-based framework designed to remove FB while preserving fine details. We propose the flicker-banding prior estimator (FPE) that predicts key banding attributes and injects it into the restoration network. Additionally, Masked Loss (ML) is proposed to concentrate supervision on banded regions without sacrificing global fidelity. To overcome data scarcity, we provide a simulation pipeline that synthesizes FB in the luminance domain with stochastic jitter in banding angle, banding spacing, and banding width. Feathered boundaries and sensor noise are also applied for a more realistic simulation. For evaluation, we collect a paired real-world FB dataset with pixel-aligned banding-free references captured via long exposure. Across quantitative metrics and visual comparisons on our real-world dataset, RIFLE consistently outperforms recent image reconstruction baselines from mild to severe flicker-banding. To the best of our knowledge, it is the first work to research the simulation and removal of FB. Our work establishes a great foundation for subsequent research in both the dataset construction and the removal model design. Our dataset and code will be released soon.
Figures & tables
Figure 1: Overview of RIFLE . The top part shows the datasets we build, including the simulated training dataset (first row) and the real-world testing dataset (second row). The bottom part shows the flicker-banding removal effect on the simulated datasets and real-world scenarios.
Figure 2: Flicker-banding (FB) when filming screens with smartphone cameras. a) Rolling shutter exposure process. b) Typical display brightness modulation (e.g., PWM and scanning refresh). c) Interaction between the camera exposure and screen modulation leads to the banding artifacts. d) Example banding patterns captured from different modern display technologies.
Figure 3: Visual comparison between our simulated flicker-banding and real-world flicker-banding. GT indicates the real-world non-banding images, on which our simulation pipeline is conducted. SIM indicates our simulation FB images, while LQ indicates real-world FB images.
Figure 4: Our flicker-banding simulation pipeline design. a) Stage 1: We generate the general framework based on the basic banding parameters and introduce parameter jitter and feather concatenation for a more realistic transition. b) Stage 2: We overlay the flicker-banding mask on the Y-channel of high-quality (HQ) images and add sensor noise to the reconstructed screen images.
Figure 5: Overview of our model design. a) We propose Banding-Suppressed High-Frequency Prior ( BSHP ), which combines a fine-grained gradient signal with a smooth structural support map to build a compact single-channel prior. b) We adopt a one-step diffusion restoration model and inject BSHP through multi-stage FiLM modulation for degradation-aware feature recalibration. c) We propose a masked loss ( ML ) that emphasizes banded regions while preserving global fidelity.
Figure 6: Visualization of BSHP. Gradient signal G captures fine transitions but also stripe responses, while repeatability map S is smoother and provides structural support. Their fusion yields a compact prior Phf that preserves informative transitions while suppressing unstable banding responses.
Methods
Task
SSIM ↑
ms-SSIM ↑
LPIPS ↓
DISTS ↓
FSIM ↑
GMSD ↓
LQ
−
0.6239
0.6758
0.3070
0.1978
0.7947
0.2045
Notch Filter
Filtering
0.5562
0.6313
0.3411
0.2600
0.7196
0.2110
MAT [ 23 ]
Super-resolution
0.4747
0.6479
0.4114
0.2243
0.7766
0.2003
InvSR [ 24 ]
Super-resolution
0.6101
0.6873
0.3833
0.2195
0.8035
0.2108
Step1X [ 25 ]
Image Editing
0.6164
0.6715
0.3564
0.2066
0.7900
0.2079
Table 1: Quantitative comparison on the real-world FB testing dataset. Except for Notch Filter and FlickerFormer*, all learning-based baselines are trained or finetuned on our simulated data. FlickerFormer* uses the official checkpoint from the original deflickering task, while FlickerFormer denotes the version finetuned on our simulated data. Best results are highlighted in red .
Figure 7: Visual comparison on three representative Real-Flicker cases. RIFLE removes stronger flicker-banding artifacts with cleaner structures and fewer residual stripes than competing baselines.
Figure 8: Visual comparison of RIFLE component ablations on a real-world FB image.
Methods
SSIM ↑
ms-SSIM ↑
LPIPS ↓
DISTS ↓
FSIM ↑
GMSD ↓
Baseline
0.6201
0.7370
0.3808
0.2048
0.8443
0.2085
ML
0.6673
0.7899
0.2520
0.1621
0.8884
0.1513
BSHP
0.6668
0.7928
0.2521
0.1587
0.8920
0.1528
ML+BSHP
0.6843
0.8318
0.2209
0.1410
0.9281
0.1376
Table 2: Ablation study of our proposed components on real-world FB datasets.
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: Flicker-banding patterns on LED matrix displays with different scanning refresh mechanisms captured under various conditions. b) Thickening of dark stripes, likely due to changes in PWM duty cycle. c-d) Different LED subpixel arrangements. e-f) Atypical banding patterns. h) Complex banding patterns caused by per-block scanning refresh.
Figure 10: Flicker-banding patterns on OLED displays with PWM driving. a) Non-uniform banding in darker areas on an iPhone OLED screen. b) Horizontal gray banding on laptop OLED. c) Extremely bending and color shifts on banding edges.
Figure 11: Flicker-banding patterns on CRT monitors with electron beam scanning. b) Phosphor persistence effect causing trailing artifacts, leaving a ghosting effect.
Figure 12: Flicker-banding patterns on DLP projectors with color wheel and micromirror switching. a-b) Comparison between original and projected images, highlighting the width of color stripes changes with brightness levels. c) Internal components adapted from DMahalko (CC BY-SA 3.0).
Figure 13: The demoiré performance comparison between no FB scene and FB scene.
Figure 14: Additional visual comparison on the remaining Section-1 cases that are not used in Fig. 7 .
Figure 15: The operating system of the User Study.
Methods
RIFLE
PiSA-SR
MAT
InvSR
Step1X
ESDNet
FlickerFormer
Notch Filter
Score/(%)
82.4
14.7
2.7
0.02
0
0.1
0.08
0
Appendix
Table 3: The results of User Study.
Figure 16: The out-of-distribution test results.
Figure 17: Per-channel visualization of banding-free images (GT) and flicker-banding images (LQ) in RGB color space. Flicker-banding artifacts appear in all three channels of the LQ images.
Figure 18: Per-channel visualization of banding-free images (GT) and flicker-banding images (LQ) in YCbCr color space. Dark flicker-banding is concentrated mainly in Y, with substantially weaker effects in the chrominance channels.
Figure 19: Visualization comparison of Y-channel switching effects. Origin shows the original image and Swapped shows the image after exchanging its Y channel with the counterpart.
Figure 20: Visual comparison of RIFLE on synthetic datasets based on RGB/ YCbCr color space.
Figure 21: A case where our method leaves residual magenta colored bands.
Figure 26: Our simulated dataset based on LSDIR [ 26 ] partial visualization (part 1).
Figure 27: Our simulated dataset based on UHDM [ 8 ] partial visualization (part 2).
Scene
Pairs
LQ exposure
LQ ISO
GT exposure
GT ISO
1
15
1/2500 or 1/500 s
360–1800
1/80 s
250
2
25
1/2500 or 1/800 s
1000–2500
1/60 s
72
3
33
1/2000 s
332–1440
1/80 s
72
4
27
1/2000 s
417–2400
1/80 s
72
5
5
1/1600 s
390–1152
1/80 s
72
Appendix
Table 4: Real-world acquisition settings for the 105 valid image pairs.
Method
Batch
Iterations
Learning rate
Peak memory
Time (h)
Notch Filter
N/A
N/A
N/A
N/A
N/A
MAT
2
45,000
2×10−4
23.1 GiB
16.2
InvSR
4
4,000
5×10−5
30.8 GiB
20.3
Step1X
8
25,000
10−4
20.5 GiB
69.0
ESDNet
16
186,875
2×10−4
26.4 GiB
50.8
FlickerFormer
1
20,000
5×10−5
17.7 GiB
6.4
Appendix
Table 5: Training settings and costs for the compared checkpoints. N/A denotes a method without training.
Method
Params (M)
Memory (GiB)
MACs (G)
FLOPs (G)
Latency (ms)
Notch Filter
0
N/A
N/A
N/A
58.4
MAT
0.690
0.4839
169
337
238.5
InvSR
983.402
3.9520
7,521
15,042
2,607.6
Step1X
12,511.620
39.9809
999,122
1,998,244
24,548.4
ESDNet
5.934
0.2238
70
140
64.1
FlickerFormer
4.155
3.7336
131
263
169.5
Appendix
Table 6: Inference costs on 512×512 inputs. M and G denote 106 and 109 . Notch filtering runs on the CPU and is not included in GPU-memory or neural-operation profiling.
Metric
LQ
PiSA-SR
RIFLE
OCR-A ↑
0.9594
0.8718
0.8841
Appendix
Table 7: OCR-based text fidelity. Higher OCR-A is better.
Method
SSIM ↑
MS-SSIM ↑
LPIPS ↓
DISTS ↓
FSIM ↑
GMSD ↓
Ewarp↓
LQ
0.6266
0.7018
0.3279
0.1994
0.7668
0.2028
17.6871
PiSA-SR
0.4921
0.6319
0.4077
0.2373
0.7086
0.2418
33.7324
RIFLE
0.6227
0.7074
0.3397
0.2031
0.7709
0.1965
20.5202
Appendix
Table 8: Full-dataset video evaluation. Ewarp values are in units of 10−3 .