PGP-DiffSR: Phase-Guided Efficient Image SR

• The paper's main contribution is the integration of progressive pruning with Fourier phase-guided restoration, reducing FLOPs and parameters while preserving structural fidelity.
• PGP-DiffSR employs a Progressive Pruning Approach to remove redundant transformer blocks and a Phase-Exchange Adapter Module to re-inject structural cues from low-quality inputs.
• Empirical results on benchmarks like DrealSR and RealSR show that the method achieves competitive perceptual quality with significantly lowered computational and memory requirements.

PGP-DiffSR is a diffusion-based image super-resolution framework that targets the computational and memory overhead of large diffusion backbones by combining architecture pruning with Fourier-phase-guided restoration. It is introduced as “Phase-Guided Progressive Pruning for Efficient Diffusion-based Image Super-Resolution” and is built around two coupled mechanisms: a Progressive Pruning Approach that removes redundant transformer blocks from a FaithDiff-style latent diffusion backbone, and a Phase-Exchange Adapter Module that re-injects structural information from the low-quality input to compensate for pruning-induced distortion (Yang et al., 2 Dec 2025).

1. Definition, scope, and nomenclature

PGP-DiffSR is defined as a lightweight diffusion method that removes redundant information from diffusion models under the guidance of the phase information of inputs for efficient image super-resolution. The method is motivated by the observation that diffusion SR systems often rely on large-scale backbones such as Stable Diffusion XL and Diffusion Transformers, leading to excessive computational and memory costs during training and inference. PGP-DiffSR therefore addresses efficiency at the backbone level rather than only accelerating samplers or adding external control modules (Yang et al., 2 Dec 2025).

The framework has two named components. The first is Phase-Guided Progressive Pruning, abbreviated in the paper as the Progressive Pruning Approach (PPA). The second is the Phase-Exchange Adapter Module (PEAM), which uses Fourier phase from the low-quality image to guide the pruned denoising network. The paper also introduces a one-step distilled variant, PGP-DiffSR-S1, built on top of the pruned and phase-guided backbone (Yang et al., 2 Dec 2025).

In this context, the acronym PGP denotes Phase-Guided Progressive Pruning. In unrelated QCSP and universal-algebra literature, PGP denotes the polynomially generated powers property (Carvalho et al., 2015, Martin, 2016). The shared acronym is terminological rather than conceptual.

2. Backbone, problem setting, and overall pipeline

PGP-DiffSR adopts FaithDiff’s diffusion model as its base backbone. The underlying system is a latent diffusion SR model with a VAE encoder/decoder, a text-conditional denoising UNet with transformer blocks in encoder, bottleneck, and decoder, and control features from low-quality images via an adapter in the spirit of T2I-Adapter or ControlNet-like conditioning. The target inefficiency is internal redundancy within this denoising UNet, including redundancy beyond the encoder and into the bottleneck and decoder (Yang et al., 2 Dec 2025).

The pipeline consists of four stages. First, the low-quality image is encoded and conditioned through the latent diffusion SR backbone. Second, PPA performs coarse and fine pruning of redundant transformer blocks. Third, PEAM injects phase-derived structural guidance into multi-scale encoder features. Fourth, the resulting pruned and phase-guided model is fine-tuned under the standard diffusion denoising objective. This pruned backbone is later reused as the teacher or base model for PGP-DiffSR-S1 (Yang et al., 2 Dec 2025).

The paper’s central empirical motivation is structural degradation under naïve pruning. It reports that phase-only reconstructions of outputs from a naïvely pruned model deviate structurally from the phase-only reconstruction of the low-quality input, indicating that pruning damages structural consistency even when textures remain plausible. The method is therefore designed not only to reduce FLOPs, parameters, and memory, but also to preserve edges, contours, and layout through phase guidance (Yang et al., 2 Dec 2025).

3. Phase-Guided Progressive Pruning

PPA begins with redundancy analysis across timesteps. For encoder, bottleneck, and decoder features, the paper computes cosine dissimilarity and KL divergence between features at consecutive timesteps. For encoder features, for example, the cosine dissimilarity is written as

$$\Phi_{\mathrm{cos}(\mathbf{E}^{(t)}_{s},\mathbf{E}^{(t+1)}_{s})} = \mathbb{E}\left[ 1 - \frac{\langle \mathbf{E}^{(t)}_{s},\,\mathbf{E}^{(t+1)}_{s}\rangle} {\|\mathbf{E}^{(t)}_{s}\|_2\,\|\mathbf{E}^{(t+1)}_{s}\|_2} \right],$$

with an analogous KL-based criterion on softmax-normalized features. The reported finding is that encoder and bottleneck features change very little across adjacent timesteps, whereas decoder features are more active on average but still contain intervals of substantial redundancy (Yang et al., 2 Dec 2025).

On the basis of this analysis, PPA applies a coarse pruning stage that reduces encoder and bottleneck modules to a single transformer block per module per scale while preserving the overall UNet topology, scales, and skip connections. After coarse pruning, the model is fine-tuned for approximately 10k iterations. The ablation reported on DrealSR gives the baseline at 2993G FLOPs and 2567M parameters, while the coarse-pruned model uses 2106G FLOPs and 1624M parameters, with quality remaining similar though somewhat degraded on DISTS (Yang et al., 2 Dec 2025).

The fine pruning stage then targets decoder redundancy at the block level. For a decoder block, the method compares input and output features using cosine dissimilarity and KL divergence. The reported pruning rule uses thresholds

$$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$

The accompanying description states that the intent is to remove blocks that change features very little. Fine pruning is performed after coarse pruning, and the smallest-scale decoder is constrained to retain four blocks (Yang et al., 2 Dec 2025).

The full pruned model, denoted “Ours” in the ablation, reaches 1555G FLOPs and 984.5M parameters. On DrealSR it reports NIQE 6.171, DISTS 0.2296, and MANIQA 0.6530, improving not only over the coarse-only stage but also over the baseline in some perceptual metrics. A control experiment with random decoder pruning at comparable compute, denoted “w/ Unfine,” produces worse DISTS, indicating that similarity-guided pruning rather than pruning alone is responsible for the final trade-off (Yang et al., 2 Dec 2025).

4. Phase-Exchange Adapter Module

PEAM is introduced to correct the structural inconsistency caused by aggressive pruning. Its premise is that Fourier phase encodes structure, while amplitude carries more detail and intensity information. The method therefore takes the phase of a low-quality adapter feature and the amplitude of a pruned UNet encoder feature, and combines them in the Fourier domain:

$$\phi^\mathbf{F}_i = \angle \mathcal{F}(\mathbf{F}_i), \qquad A^\mathbf{H}_i = \big|\mathcal{F}(\mathbf{H}_i)\big|,$$

$$\tilde{\mathbf{H}}_i = \mathcal{F}^{-1}\left( A^\mathbf{H}_i e^{j \phi^\mathbf{F}_i} \right).$$

Here, $\mathbf{F}_i$ is a low-quality feature extracted by the adapter, and $\mathbf{H}_i$ is the corresponding encoder feature from the pruned denoising UNet (Yang et al., 2 Dec 2025).

The phase-exchanged feature is not inserted as a hard replacement. Instead, PEAM concatenates $\mathbf{H}_i$ and $\tilde{\mathbf{H}}_i$, applies a shallow fusion block consisting of a $3\times3$ convolution and LeakyReLU, and then applies eleven stacked $3\times3$ convolutional layers for deeper refinement. The refined feature is split channel-wise into two parts and used in a FiLM-like modulation:

$$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$0

The resulting $$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$1 replaces the original encoder feature for subsequent UNet computation. PEAM is inserted at multiple encoder levels, so phase guidance is multi-scale rather than single-scale (Yang et al., 2 Dec 2025).

The ablation on DrealSR isolates the value of phase exchange from the value of added capacity. Simple feature addition (“w/ Add”) gives NIQE 7.070, DISTS 0.2483, and MANIQA 0.6476. Direct phase replacement without learnable fusion (“w/ Phase change”) improves NIQE to 6.296 and DISTS to 0.2402. A same-capacity convolutional fusion block without the phase exchange mechanism (“w/ Fusion conv”) yields NIQE 6.638 and DISTS 0.2518. Full PEAM reaches the best values among these variants: NIQE 6.171, DISTS 0.2296, and MANIQA 0.6530. The paper therefore attributes the gain specifically to phase exchange combined with learnable modulation, not merely to auxiliary convolutions (Yang et al., 2 Dec 2025).

5. Training procedure, distilled variant, and efficiency profile

PGP-DiffSR retains the standard latent diffusion denoising objective,

$$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$2

where the novelty lies in the architecture of $$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$3, not in a new diffusion loss. Training proceeds progressively: the model starts from the baseline FaithDiff-like system, undergoes coarse pruning and fine-tuning, then fine pruning and another fine-tuning stage, with PEAM trained jointly during this process. Reported settings include patch size $$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$4, batch size 128 for the multi-step model, AdamW with learning rate $$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$5 and cosine annealing, Euler sampling with 20 steps, and classifier-free guidance scale 5 (Yang et al., 2 Dec 2025).

PGP-DiffSR-S1 is the one-step distilled extension. It uses PGP-DiffSR as teacher or base and adds a DAPE model for prompt extraction. Its training is two-stage: 4k iterations with MSE + LPIPS loss, followed by 14k iterations with LPIPS + Classifier Score Distillation loss. The optimizer is Adam with learning rate $$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$6; trainable LoRA modules use rank 16; batch size is 16 and patch size remains $$\Phi_{\mathrm{cos}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 0.2 \quad\text{and}\quad \Psi_{\mathrm{KL}(\mathbf{D}_s^{(\mathrm{In})},\mathbf{D}_s^{(\mathrm{Out})})} \geq 1.0.$$7 (Yang et al., 2 Dec 2025).

The reported efficiency gains are substantial. For 1024×1024 patches on an A800 GPU, the full FaithDiff-like baseline uses 2993G FLOPs and 2567M parameters, while PGP-DiffSR uses 1555G FLOPs and 984.5M parameters. In runtime terms, with 20 sampling steps, FaithDiff requires 2.55 s and 10.40 GB GPU memory, whereas PGP-DiffSR requires 0.78 s and 5.69 GB (Yang et al., 2 Dec 2025).

Configuration	Compute footprint	Runtime footprint
FaithDiff-like baseline	2993G FLOPs, 2567M params	2.55 s, 10.40 GB
PGP-DiffSR	1555G FLOPs, 984.5M params	0.78 s, 5.69 GB

On benchmark datasets, the paper emphasizes no-reference perceptual quality. On RealSR, PGP-DiffSR reports CLIP-IQA+ 0.6056, MUSIQ 70.17, and MANIQA 0.6841, all listed as best, with NIQE 5.467 among the best. On DrealSR it reports NIQE 6.171, CLIP-IQA+ 0.6097, MUSIQ 67.50, and MANIQA 0.6530, each listed as best. On RealPhoto60 it remains competitive, with CLIP-IQA+ 0.6166 and MUSIQ 72.81. The one-step PGP-DiffSR-S1 similarly leads other one-step SR methods on no-reference perceptual metrics across RealSR, DrealSR, and RealPhoto60, while remaining less dominant on LPIPS and DISTS (Yang et al., 2 Dec 2025).

6. Relation to diffusion SR literature, limitations, and significance

PGP-DiffSR belongs to a different branch of diffusion SR research than methods that repurpose a fixed pretrained generator without retraining. For example, Diff-SR studies arbitrary-scale super-resolution by controlling the noise level and reverse-diffusion starting point of a pretrained diffusion generative model, without fine-tuning or distillation (Li et al., 2023). By contrast, PGP-DiffSR starts from a FaithDiff-style latent diffusion SR system, prunes the denoising backbone itself, and restores lost structure through explicit phase guidance (Yang et al., 2 Dec 2025).

The method is also differentiated from prior efficiency work in the paper’s own framing. Control-based SR approaches lighten training but still deploy a full backbone at inference, distillation requires a large teacher during training, and sampler acceleration reduces step count without shrinking backbone complexity. PGP-DiffSR instead reduces the model internally across encoder, bottleneck, and decoder, and augments the pruned model with PEAM. The authors further state that, to their knowledge, it is the first method to explicitly use Fourier phase from low-quality images as guidance for a diffusion-based SR model (Yang et al., 2 Dec 2025).

Several limitations are presented in the supplied analysis as implications rather than as an explicit limitations section in the paper. This suggests that the framework is more optimized for perceptual quality than for strict pixel-wise fidelity, since GAN-based methods can still obtain lower LPIPS or DISTS in some settings. A plausible implication is that PEAM may be sensitive when low-quality phase itself is heavily corrupted, for example by extreme motion blur or very strong noise. Another plausible implication is that behavior at larger scaling factors remains unresolved, since the reported experiments focus on real-world ×2 SR rather than ×4 or ×8. The pruning thresholds are also reported as empirically chosen, so transfer to other backbones or restoration tasks would require re-validation (Yang et al., 2 Dec 2025).

Within diffusion SR, PGP-DiffSR is significant chiefly as an efficiency-oriented architecture design. Its contribution is not a new generative objective or a new sampler, but a coupled compression-and-correction strategy: feature-similarity-guided backbone pruning plus phase-guided structural anchoring. The resulting system occupies a distinct position between large-capacity latent diffusion SR backbones and purely inference-side acceleration methods, while retaining competitive restoration quality on no-reference perceptual metrics and enabling a one-step distilled derivative (Yang et al., 2 Dec 2025).