Flicker-Banding Prior Estimator (FPE)

• Flicker-Banding Prior Estimator (FPE) is a restoration component that infers structured banding artifacts from degraded images or videos using parametric or dense field representations.
• It employs either a ResNet-based module for global parameter regression or a Continuous Prior Perception module via Video-SwinUNet for spatiotemporal amplitude mapping.
• By conditioning restoration networks with explicit artifact cues, FPE improves band removal performance and achieves superior results on both synthetic and real datasets.

A Flicker-Banding Prior Estimator (FPE) is a restoration component that predicts structured priors of flicker-banding from a degraded image or video and supplies those priors to a downstream reconstruction model. In the recent literature on screen-capture restoration, the term is used explicitly in RIFLE for an auxiliary network that regresses banding attributes from a flicker-banded image (Zhu et al., 29 Sep 2025), while VDFP realizes the same functional role through a Continuous Prior Perception (CPP) module that estimates a continuous spatial-temporal flicker amplitude map for video deflickering (Zhou et al., 20 May 2026). In both cases, the estimated prior is not the restored signal itself; it is an intermediate representation that constrains where banding exists, how strong it is, and, in the video setting, how it evolves across time.

1. Definition and artifact model

Flicker-banding arises when a smartphone with a rolling-shutter CMOS sensor records a digital display whose luminance is modulated periodically in time. The aliasing between display modulation and line-by-line camera sampling produces alternating bright and dark stripes that may be curved, cracked, diamond-shaped, or organized at multiple spatial scales (Zhou et al., 20 May 2026). RIFLE defines the same phenomenon as alternating bright–dark stripes caused by temporal aliasing between a camera’s rolling-shutter readout and the display’s brightness modulation (Zhu et al., 29 Sep 2025).

Within this setting, a “prior” is a representation of the banding structure before restoration. Two forms appear in the literature. One is parametric: orientation angle $\theta$, stripe width $w$, stripe spacing or gap $g$, and associated jitter parameters define a family of stripe layouts in a stripe-aligned coordinate system (Zhu et al., 29 Sep 2025). The other is field-based: a spatiotemporal degradation field $O(x,y,t)$ or a continuous confidence map $\hat{A}_{\text{flicker}}\in[0,1]$ describes the local presence and amplitude of banding over space and time (Zhou et al., 20 May 2026).

This distinction is consequential. A parametric prior expresses the artifact through a low-dimensional physical or morphological description. A field prior expresses it as a dense spatial or spatiotemporal signal. This suggests that “FPE” is better understood as a family of estimators whose common purpose is to infer artifact structure from degraded observations, rather than as a single canonical architecture.

2. Estimation paradigms

The two main instantiations in current work differ in whether they estimate global attributes or dense prior fields.

System	Estimated prior	Estimator
RIFLE	$\hat{\theta}$, $\hat{w}$, $\hat{g}$	ResNet-based FPE
VDFP	$$\hat{A}_{\text{flicker}}\in[0,1]^{T\times H\times W}$$	CPP with Video-SwinUNet

In RIFLE, the FPE is an auxiliary network that takes a flicker-banded image as input and predicts explicit physical or morphological attributes of the banding pattern. The selected parameters named in the paper are banding width $W$, banding spacing $w$0, and banding angle $w$1. The estimator is described as “inverse to the FB simulation process”: the simulator generates banding from known parameters, and the FPE learns to recover those parameters from the simulated low-quality image (Zhu et al., 29 Sep 2025). The prior is therefore vector-valued and global.

In VDFP, the effective FPE is the CPP module, a first-stage Video-SwinUNet that maps a flickered video $w$2 to a continuous confidence or amplitude map,

$w$3

The output is a one-channel confidence field with the same spatio-temporal resolution as the input sequence. Unlike binary segmentation, CPP is trained to represent gradual luminance transitions, band edges, and temporal phase trajectories (Zhou et al., 20 May 2026). The prior is therefore dense and spatiotemporal.

The literature is explicit that these roles must not be conflated with the synthetic generators used during training. In VDFP, Degradation Field Modeling (DFM) is a forward generator of banding fields from parameters and is not an estimator at inference time; CPP is the estimator. This clarification is central to the concept of FPE in video deflickering (Zhou et al., 20 May 2026).

3. Physical modeling, supervision, and synthetic data

Both major formulations depend on synthetic data because paired real training data are scarce. In both cases, the simulation is not merely data augmentation; it defines the supervisory signal that the prior estimator learns to invert.

RIFLE constructs a stripe-aligned parametric model in rotated coordinates $w$4, with stripe period

$w$5

The synthetic banding mask $w$6 is then perturbed by orientation angle jitter, spacing jitter, width jitter, and axial edge fluctuation. Banding is applied only on luminance after RGB-to-YCbCr conversion, and the final simulated low-quality image includes sensor noise (Zhu et al., 29 Sep 2025). This provides exact ground-truth banding parameters $w$7, $w$8, $w$9, and a derived mask $g$0 for each synthetic pair.

VDFP formulates a rolling-shutter-based spatiotemporal degradation field. Pixel position is projected onto the axis orthogonal to stripe direction; temporal phase shift is added through motion parameters $g$1; and a soft occupancy field is generated by a smoothstep-based stripe model. Base high-frequency stripes and thick low-frequency stripes are simulated independently and fused as

$g$2

From this field, the method derives the luminance-fluctuation amplitude map $g$3 used to supervise CPP (Zhou et al., 20 May 2026).

The training losses reflect these different targets. VDFP uses Flicker-Aware MSE (FA-MSE),

$g$4

which is mathematically an MSE but is defined on a continuous luminance-fluctuation amplitude target rather than on binary labels (Zhou et al., 20 May 2026). RIFLE pretrains FPE as a scalar-parameter regressor; the paper does not provide a LaTeX loss formula for FPE, but it states that the estimator predicts the selected parameters inverse to the simulation process and is pretrained before integration into the restoration network (Zhu et al., 29 Sep 2025).

The corresponding evaluation datasets also differ. VDFP introduces DeViD, a real-world dataset of 108 videos: 60 from SPMCS, YouHQ, and UDM10, and 48 newly captured from real screens, covering stages, red carpets, sports, ads, video games, cartoons, movies, and TV series (Zhou et al., 20 May 2026). RIFLE reports a paired real-world FB dataset with pixel-aligned banding-free references captured via long exposure (Zhu et al., 29 Sep 2025).

4. Conditioning restoration models

The operational purpose of an FPE is to condition a powerful restoration backbone with explicit artifact structure. The conditioning strategy differs sharply between RIFLE and VDFP, but both are designed to reduce ambiguity between banding and legitimate image content.

RIFLE is built on PiSA-SR, a one-step latent diffusion restoration model with a pre-trained VAE encoder/decoder and a UNet backbone. The pre-trained FPE is introduced through an embedding module that maps the predicted parameter vector $g$5 into a conditioning embedding for the UNet. During RIFLE training, the VAE and FPE are frozen, the UNet is fine-tuned with LoRA, and the embedding module is trained with full parameters (Zhu et al., 29 Sep 2025). The prior here is a compact attribute vector: it conveys approximate stripe angle, width, and spacing rather than a dense spatial mask.

VDFP instead injects a dense prior channel into a video diffusion backbone derived from STAR. The degraded latent feature $g$6 originally has four channels with first-layer weights $g$7. To condition on the CPP prior, VDFP concatenates the estimated map as a fifth channel and expands the first-layer weights to

$g$8

Initially,

$g$9

so the model behaves exactly like the original pre-trained model; only during fine-tuning does the new channel begin to influence inference (Zhou et al., 20 May 2026). This “zero-initialized augmented input layer” is the explicit mechanism by which the learned flicker prior is injected without immediately disturbing the pre-trained text-to-video prior. The pre-trained T2V backbone is frozen, and Video-ControlNet plus Local Information Enhancement Modules (LIEM) are fine-tuned, as in STAR (Zhou et al., 20 May 2026).

These two integration patterns correspond to two different understandings of prior information. RIFLE treats the prior as a low-dimensional descriptor of the degradation family. VDFP treats the prior as a dense conditioning field that functions as a soft spatial-temporal control signal. A plausible implication is that the choice between them depends on whether the artifact can be summarized adequately by a few global parameters or requires local phase and amplitude information.

5. Empirical evidence

The empirical record supports the claim that explicit prior estimation improves restoration over relying on implicit inference inside the main network.

In RIFLE’s simulated-data ablation, the baseline achieves PSNR 22.12, SSIM 0.7490, ms-SSIM 0.8677, LPIPS 0.1812, DISTS 0.0920, FSIM 0.9406, and GMSD 0.1297. Adding FPE alone yields PSNR 22.15, ms-SSIM 0.8683, DISTS 0.0918, FSIM 0.9431, and GMSD 0.1337. The full ML + FPE model yields PSNR 22.30, ms-SSIM 0.8697, DISTS 0.0908, FSIM 0.9460, and GMSD 0.1286, and is reported as best or second-best across all metrics. The qualitative interpretation in the paper is that FPE introduces an explicit prior that suppresses periodic stripes more aggressively, while ML and FPE together produce the cleanest banding removal with minimal residuals while preserving edges and fine details (Zhu et al., 29 Sep 2025).

VDFP reports that its main quantitative table on DeViD shows improvement over all baselines on SSIM, LPIPS, VMAF, DISTS, MUSIQ, DOVER, and temporal warping error. The paper also presents Y–T slice analysis showing that the predicted prior $O(x,y,t)$0 follows the diagonal trajectory of moving banding, and temporal curves in which peaks in the predicted confidence align with valleys in luminance. Its ablation study indicates that STAR without DFM+CPP improves low-quality input only slightly, DFM training significantly improves restoration, and adding CPP yields the best metrics and removes residual artifacts left by DFM-only training (Zhou et al., 20 May 2026).

These findings support two narrower conclusions. First, explicit prior estimation is most beneficial when banding is dominant and structured rather than weak and stochastic. Second, a prior estimator is particularly valuable when the main restoration network is a large diffusion model, because the estimator reduces the burden of disentangling artifact geometry from semantic content.

6. Related perceptual models, limitations, and nomenclature

FPE design is closely related to perceptual models of flicker visibility, even when those models are not themselves restoration systems. An eccentricity-dependent spatio-temporal flicker fusion model predicts that temporal modulation is imperceptible when

$O(x,y,t)$1

with dependence on eccentricity $O(x,y,t)$2, spatial frequency $O(x,y,t)$3, and luminance $O(x,y,t)$4 (Krajancich et al., 2021). The elaTCSF framework extends temporal contrast sensitivity by incorporating luminance, eccentricity, and area, and defines an effective sensitivity $O(x,y,t)$5 together with a flicker index

$O(x,y,t)$6

for applications such as variable refresh rate flicker and low-persistence VR displays (Cai et al., 21 Mar 2025). These results suggest that future FPEs could incorporate not only artifact geometry but also local visibility thresholds.

A complementary line of work is no-reference spatial banding detection. FS-BAND models banding as “high-frequency false edges contained in low-frequency smoothing regions,” using a high-frequency map from isotropic Sobel filtering, a low-frequency map from a Mumford–Shah piece-wise smooth approximation, and a dual-branch ResNet-50 to generate patch-level banding labels and a pixel-wise banding map (Chen et al., 2023). FS-BAND is purely spatial and does not estimate flicker-banding in video, but its decomposition of banding into high-frequency false contours on low-frequency backgrounds is directly relevant to FPE design.

Current FPE formulations also have clear limitations. RIFLE’s FPE is a global parameter regressor, so locally varying angles or multiple overlapping banding frequencies may not be captured well by a single parameter set; the paper also notes that FPE alone can at times slightly soften textures (Zhu et al., 29 Sep 2025). VDFP trains its prior estimator on synthetic data generated by DFM and validates on real DeViD videos, but the described real-world hardware patterns are still constrained by the modeled family, and the summary notes that DeViD mainly contains LED matrix banding (Zhou et al., 20 May 2026). In both cases, synthetic supervision is central, which improves controllability but leaves open the question of generalization to unusual display driving schemes.

Finally, the acronym is ambiguous across fields. In robust covariance estimation, FPE denotes Tyler’s Fixed-Point Estimator rather than Flicker-Banding Prior Estimator (Pascal et al., 2013). The two usages are unrelated except at the level of acronym collision. In image and video restoration, “FPE” refers to an estimator of structured banding priors; in robust statistics, it refers to a fixed-point estimator of scatter matrices.

Across these formulations, the common principle is stable: an FPE estimates artifact structure that is easier to supervise in a synthetic or physics-informed space than to recover implicitly during restoration. Whether expressed as $O(x,y,t)$7 or as $O(x,y,t)$8, the prior serves as an intermediate representation that localizes banding, quantifies its strength, and constrains how aggressively the downstream model should modify the signal.