.png){kind=icon}
.png){kind=icon}
Return
Beyond deterministic scoring: Probabilistic quality assessment of generative faces via risk-aware correspondence
.png){kind=icon}
PRE
AI
Z
L
S
L
Z
DOI：10.1016/j.displa.2026.103433.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As generative face restoration and synthesis are deployed in high-fidelity applications, quality assessment faces novel challenges, including 'hallucinated' identity shifts and spectrally localized artifacts that evade traditional spatial metrics. Existing Full-Reference (FR) metrics, predominantly based on deterministic Transformer backbones, suffer from two critical limitations: a spatial-modality bias that underweights high-frequency generative noise and a lack of uncertainty calibration regarding correspondence reliability. To bridge this gap, FGMH-IQA is proposed as a probabilistic framework that reframes facial quality assessment from a point estimation to a risk-aware decision process. First, this paper presents a Spatial Dependency Metric (SDM) that models patch correspondences as multivariate Gaussian latents, capturing epistemic uncertainty to explicitly down-weight unreliable alignments in identity-critical regions. Second, a Multispectral Visual Compensation (MVC) backbone is devised using Haar wavelets to counteract the low-pass tendency of Transformers; this module functions as a spectral equalizer, injecting multi-resolution high-frequency cues to detect subtle generative artifacts. Finally, an uncertainty-gated Adaptive Evaluator is incorporated to dynamically balance fidelity and naturalness, enabling robust scoring even under imperfect reference scenarios common in restoration pipelines. Extensive experiments on FIQA and PIPAL demonstrate that FGMH-IQA aligns closely with human visual psychophysics, surpassing state-of-the-art baselines by explicitly quantifying the confidence of its quality judgments.
Keywords:
Generative face assessment
Epistemic uncertainty
Spectral equalizer
Human visual system (HVS)
Risk-aware IQA
Journal
IF:
3.4
Papers:
2.1K
Citations:
3.2K
Organization
Cited Papers