Panini-Net: Degradation-Aware Face Restoration

• Panini-Net is a degradation-aware GAN prior network that fuses degraded-image features with StyleGAN2 features using unsupervised degradation representation learning for face restoration and 16× super-resolution.
• It employs a multi-scale architecture with an Image Feature Extraction module, latent code encoder, and a novel degradation-aware feature interpolation (DAFI) mechanism to balance realness and fidelity under blur, noise, and compression.
• Empirical results demonstrate state-of-the-art performance in multi-degradation face restoration, showing improved PSNR, FID, and LPIPS metrics while preserving identity and fine details.

Searching arXiv for relevant papers on "Panini-Net" and disambiguating exact usage.

Searching arXiv for: "Panini-Net GAN Prior Based Degradation-Aware Feature Interpolation for Face Restoration"

Panini-Net is a GAN Prior based degradation-aware feature interpolation network for face restoration (FR) that was introduced to address the realness–fidelity trade-off under multi-degradation settings such as blur, downsampling, noise, and compression (Wang et al., 2022). The method uses a pre-trained StyleGAN2 generator as GAN prior, explicitly learns degradation representations, and conditions multi-scale feature fusion on those representations through a degradation-aware feature interpolation mechanism. In the formulation reported for the method, Panini-Net targets both multi-degradation face restoration and face super-resolution, with an unsupervised degradation representation learning strategy, progressive fusion of degraded-input features and GAN-prior features, and empirical results showing state-of-the-art performance for multi-degradation FR and competitive performance for 16× face SR (Wang et al., 2022).

1. Problem setting and design objective

Face restoration is treated as an ill-posed inverse problem, and the difficulty increases under multi-degradation settings that combine blur, downsampling, noise, and compression (Wang et al., 2022). The paper situates Panini-Net in the line of FR methods that use pre-trained generative adversarial networks, specifically StyleGAN2, as GAN prior. These methods improve perceptual quality, or realness, but are reported to struggle to balance realness and structural fidelity across varying degradation levels. Latent-space inversion approaches are described as having limited spatial expressivity and often causing identity inconsistency, while hybrid methods that combine image features with GAN-prior features are described as lacking explicit, degradation-aware fusion (Wang et al., 2022).

Panini-Net is designed around the claim that restoration quality depends not only on access to a strong generative prior but also on modulating how much the network trusts degraded-image evidence versus prior-generated features at different degradation levels. The method therefore learns a degradation representation, denoted DR, and uses it to adaptively fuse two informative feature streams: features extracted from the degraded input image and features extracted from the StyleGAN2 prior (Wang et al., 2022). This suggests that Panini-Net should be understood less as a pure inversion method and more as a degradation-conditioned feature-routing architecture built on top of a fixed generative backbone.

The stated objective is to improve robustness and the realness–fidelity trade-off across diverse degradation levels in both multi-degradation FR and face super-resolution (Wang et al., 2022). The method also aims to reduce the visual quality gap between restored outputs and images directly generated by the pre-trained GAN prior.

2. Core architecture

Panini-Net comprises four principal components: an Image Feature Extraction (IFE) module, a degradation representation encoder trained by Unsupervised Degradation Representation Learning (UDRL), a Latent Code Encoder (LCE), and a Degradation-Aware Feature Interpolation (DAFI) mechanism integrated with a pre-trained StyleGAN2 generator (Wang et al., 2022).

The IFE module extracts multi-scale features from the degraded input. Its progressive feature extraction is defined as

$\mathbf{\tilde{F}_{IFE}^{i} = \begin{cases}\mathcal{H}_{IFE}^{i}(\mathbf{X}), & i = 6;\\mathcal{H}_{IFE}^{i}(\mathbf{\tilde{F}_{IFE}^{i+1}), & 1\le i<6,\end{cases}$

and the decoupling branch produces the final fusion features

$\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$

The same encoder stack also predicts an extended latent code for the GAN prior,

$\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$

The GAN Prior Module (GPM) is a pre-trained StyleGAN2 generator. It starts from a learned constant $\mathbf{F}_{init}$ and generates progressive features through GAN Prior Blocks (GPB). The forward path is specified as

$$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$

with final image synthesis

$$\mathbf{Y} = \mathcal{H}_{GPB}^{9}(\mathbf{F}^{8}_{GPB}).$$

A key architectural decision is that only the early GPBs are fused with DAFI, specifically five DAFI blocks, while later GPBs remain untouched (Wang et al., 2022). The paper motivates this as a way to correct coarse structures and identity in early layers while preserving StyleGAN2’s fine-texture quality in higher layers. No explicit spatial warping or alignment is introduced.

All related features $\mathbf{\tilde{F}_{IFE}^{i}}$, $\mathbf{F}_{IFE}^{i}$, $\mathbf{F}_{GPB}^{i}$, and $\mathbf{F}_{DAFI}^{i}$ are specified to share the same shape $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$0 at each stage (Wang et al., 2022). This alignment makes the fusion operator an interpolation rather than a concatenation-based fusion requiring additional spatial reconciliation.

3. Degradation representation learning and DAFI

The degradation representation is a global embedding vector intended to summarize degradation mode independently of image content, covering blur, downsampling, noise, and compression (Wang et al., 2022). Panini-Net learns this representation through UDRL, a contrastive pre-training strategy that follows a MoCo-style framework with an Encoder and a Momentum Encoder.

For each training step, the same degradation function is applied to two different high-quality images to form a positive pair. Let $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$1 and $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$2 denote the degraded images produced in this way. The Encoder produces $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$3 from $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$4, the Momentum Encoder produces $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$5 from $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$6, and a queue provides negatives $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$7. The loss is

$\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$8

This construction is explicitly intended to discourage content matching and to focus the embedding on degradation mode (Wang et al., 2022). After pre-training, the fixed DRE maps an input $\mathbf{F}_{IFE}^{i} = \mathcal{H}_{Conv}^{i}(\mathbf{\tilde{F}_{IFE}^{i}),\quad i\in\{1, ..., 5\}.$9 to a 256-D vector,

$\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$0

DAFI is the mechanism that operationalizes this degradation signal. At each fusion stage $\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$1, an MLP processes $\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$2 and a Softmax yields a channel-wise interpolation mask:

$\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$3

The fused feature is then

$\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$4

where $\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$5 denotes per-channel multiplication broadcast spatially (Wang et al., 2022). The mask is therefore channel-wise rather than spatially adaptive. The paper describes this as lightweight and parameter-efficient relative to concatenation-plus-convolution fusion.

The interpolation is progressive across multiple scales. Earlier DAFI blocks affect coarse structures, and later blocks have stronger impact, with DAFI5 specifically used for analysis of interpolation ratios (Wang et al., 2022). Because the interpolation weights are deterministic functions of DR, adjusting DR or directly biasing the mask entries enables editable restoration. The authors report that both masked components, $\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$6 and $\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$7, can each generate reasonable outputs, which they interpret as evidence of practical editability.

4. Training protocol, degradation model, and variants

Training proceeds in two stages (Wang et al., 2022). First, UDRL is used to pre-train the DRE with the contrastive loss $\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$8. Second, the pre-trained StyleGAN2 generator is loaded, the DRE parameters are fixed, and the remaining modules are fine-tuned.

The fine-tuning losses reported in the main text are a standard L1 reconstruction loss, a VGG-based perceptual loss, and a vanilla GAN adversarial loss using the pre-trained StyleGAN2 discriminator (Wang et al., 2022). The paper states that explicit equations or weights for these losses are not provided in the main text and are referenced to the appendix.

For multi-degradation FR, the degradation model is

$\mathbf{w}^{+} = \mathcal{H}_{LCE}(\mathbf{\tilde{F}_{IFE}^{1}), \quad \mathbf{w}^{+}\in\mathbb{R}^{18\times512}.$9

This includes Gaussian blur with size $\mathbf{F}_{init}$0, downsampling with rate $\mathbf{F}_{init}$1, additive noise with strength $\mathbf{F}_{init}$2, and JPEG compression with quality factor $\mathbf{F}_{init}$3 (Wang et al., 2022). The reported parameter ranges are $\mathbf{F}_{init}$4, $\mathbf{F}_{init}$5, and $\mathbf{F}_{init}$6, with $\mathbf{F}_{init}$7 randomly selected per high-quality image.

The reported training setup for multi-degradation FR uses paired HQ–LQ images synthesized from CelebA-HQ or FFHQ depending on the task, with LQ/HQ sizes of 512×512 and 1024×1024, Adam optimization, Cosine Annealing scheduling, batch size 8, and 600K iterations (Wang et al., 2022).

A separate 16× face SR variant is also described. In that setting, the DRE is removed and $\mathbf{F}_{init}$8 becomes a learnable constant vector, while input/output sizes are adjusted for 64×64 low-resolution input and 1024×1024 high-resolution output with bilinear downsampling used to construct training pairs (Wang et al., 2022). This variant implies that in a single, fixed degradation regime, explicit sample-wise degradation encoding is less central than in multi-degradation FR.

5. Empirical results, ablations, and behavior under degradation severity

Panini-Net is evaluated on two tasks: multi-degradation FR and 16× face SR (Wang et al., 2022). In multi-degradation FR on CelebA-HQ, compared against GFP-GAN and GPEN, the reported quantitative results are: GFP-GAN with PSNR 17.22, FID 34.61, LPIPS 0.4150; GPEN with PSNR 17.91, FID 32.03, LPIPS 0.4355; and Panini-Net with PSNR 18.01, FID 24.66, LPIPS 0.4470 (Wang et al., 2022). The paper describes this as state-of-the-art performance for multi-degradation face restoration.

In 16× face SR on CelebA-HQ, compared against pSp and GLEAN, the reported results are: pSp with PSNR 12.90, FID 47.65, LPIPS 0.6529; GLEAN with PSNR 21.66, FID 19.76, LPIPS 0.4013; and Panini-Net with PSNR 21.19, FID 16.77, LPIPS 0.3886 (Wang et al., 2022). These numbers show that Panini-Net does not maximize PSNR in that setting, but it achieves the best FID and LPIPS among the listed methods. A plausible implication is that the method prioritizes perceptual realism and feature quality over strict pixelwise fidelity in the super-resolution regime.

The qualitative findings reported for multi-degradation FR are that Panini-Net yields the best balance between realness and identity consistency, especially under severe degradations where GFP-GAN deteriorates and GPEN lacks realistic detail (Wang et al., 2022). For 16× SR, the paper states that pSp achieves high visual quality but poor identity preservation, GLEAN is strong, and Panini-Net is competitive or superior in fine details such as eyes and teeth, which the authors attribute to DAFI’s effective fusion.

The ablation study directly compares DAFI against concatenation-plus-convolution fusion. In 16× SR, Panini-Net with Cat-Conv reports PSNR 21.17 and FID 18.41, whereas Panini-Net with DAFI reports PSNR 21.19 and FID 16.77 (Wang et al., 2022). The interpretation given is that interpolation preserves GAN-prior details better than Cat-Conv and that DR-conditioned fusion improves robustness across degradations.

The behavior of the interpolation masks under increasing degradation severity is analyzed using the DAFI5 usage ratio

$\mathbf{F}_{init}$9

With fixed $$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$0, $$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$1, and $$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$2, and varying downsampling rate $$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$3, the reported values are $$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$4, $$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$5, $$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$6, and $$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$7, respectively (Wang et al., 2022). The paper interprets this monotonic increase as evidence that Panini-Net increases reliance on GAN-prior features as degradation becomes more severe. This is consistent with the intended role of DR-conditioned interpolation.

The same section also reports masked-generation dissections. Using only input features corresponds to

$$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$8

and using only GAN-prior features corresponds to

$$\mathbf{F}_{GPB}^{i} = \begin{cases}\mathcal{H}_{GPB}^{i}(\mathbf{F}_{init}), & i = 1\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{DAFI}), & 1<i<7\\mathcal{H}_{GPB}^{i}(\mathbf{F}^{i-1}_{GPB}). & 7 \leq i<9\end{cases}$$9

The reported observation is that with mild degradation, input features contribute coarse structure while the GAN prior contributes refined details, whereas with severe degradation the input-valid content declines and the GAN prior fills missing structure (Wang et al., 2022).

6. Limitations, scope boundaries, and terminological ambiguity

The paper identifies several limitations for the face-restoration Panini-Net (Wang et al., 2022). Extreme cases such as severe occlusions, atypical poses, lighting, and domain shifts may still challenge identity consistency and may induce hallucinations because of the reliance on GAN priors. The current DAFI uses channel-wise masks, and spatially adaptive fusion is not explicitly realized. Future directions named in the paper include alternative interpolation forms such as spatial-wise interpolation, new mask-generation strategies, and extensions of editable FR through mask biasing or conditioning DR for multiple plausible restorations.

The name “Panini-Net” is not unique across arXiv. In a different and unrelated line of work, “Panini-Net” denotes the network-based formalization of Pāṇini’s derivational system for automating Sanskrit derivative noun formation, where the “net” is a directed acyclic graph of rule classes with inheritance and notification edges rather than a neural network for image restoration (Krishna et al., 2015). Separately, “Panini” has also been used for a non-parametric continual learning framework in token space based on Generative Semantic Workspaces and structured memory traversal for question answering (Rajesh et al., 16 Feb 2026). These usages are terminologically adjacent but technically independent.

Within the computer-vision literature, however, Panini-Net specifically refers to the degradation-aware GAN-prior interpolation architecture introduced for face restoration (Wang et al., 2022). Its central technical contribution is the combination of unsupervised degradation representation learning with channel-wise, multi-scale interpolation between degraded-image features and StyleGAN2 prior features, yielding a restoration pipeline that is explicitly conditioned on degradation severity rather than relying on fixed or implicit fusion rules.