One-step diffusion-based real-world image super-resolution with visual perception distillation

Xue Wu, Jingwei Xin, Jun Hao, Hui Gao, Jie Li, Nannan Wang, Xinbo GaoNeurocomputing 2026 · DOI

TL;DR

This paper introduces VPD-SR, a novel one-step diffusion model for real-world image super-resolution (SR). It tackles the slow inference of diffusion models by distilling a powerful teacher model into a single-step student model, enhanced with explicit semantic supervision from CLIP and a high-frequency perception loss to improve visual quality.

What Problem Does This Paper Address?

Image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) counterparts, with applications ranging from medical imaging to satellite imagery. While diffusion models have shown impressive results in generating perceptually high-quality images due to their ability to model complex data distributions, they typically require dozens or even hundreds of sampling steps for inference. This makes them impractical for real-time applications. Existing methods to accelerate diffusion-based SR, such as knowledge distillation, often struggle to maintain semantic alignment and perceptual quality, leading to suboptimal results. This paper addresses the challenge of achieving both high perceptual quality and semantic fidelity in a single-step diffusion-based SR model.

Methods

The core of VPD-SR is a one-step diffusion distillation framework. Instead of training a diffusion model from scratch or using complex multi-step sampling, they distill knowledge from a pre-trained, multi-step "teacher" model (specifically, ResShift) into a "student" model that can generate an HR image in a single step.

To overcome the limitations of standard distillation, VPD-SR introduces two key novel components:

  • Explicit Semantic-aware Supervision (ESS): This leverages the powerful visual-semantic understanding capabilities of the pre-trained CLIP model. By extracting semantic embeddings from both the ground truth (GT) and the student model's output, they enforce semantic alignment in the latent space. This helps ensure that the generated image not only looks good but also accurately reflects the content of the original image.
  • High-Frequency Perception (HFP) Loss: The authors observe that high-frequency details are crucial for visual perception. In addition to a standard distillation loss, the HFP loss specifically guides the student model to restore missing high-frequency details in degraded images. This is achieved by analyzing the high-frequency components of the teacher and student model outputs using Discrete Wavelet Transform (DWT).

Additionally, they incorporate adversarial training using a patch-based discriminator to further enhance the authenticity and realism of the generated images.

Datasets

The paper uses the following datasets:

  • Training: ImageNet [54] (synthetically degraded).
  • Evaluation:
    • Synthetic: ImageNet-Test [23]
    • Real-world: RealSR [55], RealSet65 [23], DRealSR [56]

Key Insights and Ideas

  • High-frequency details are critical for perceptual quality: The analysis of teacher model outputs at different sampling steps reveals that high-frequency components strengthen significantly in later stages, correlating with improved perceptual quality metrics like CLIPIQA. This insight directly motivates the HFP loss.
  • CLIP's semantic understanding is a powerful tool for SR: Using CLIP's pre-trained cross-modal alignment capabilities provides explicit semantic guidance, which is crucial for ensuring semantic consistency and preventing semantic ambiguity in the generated images, especially from severely degraded inputs.
  • One-step distillation can achieve SOTA performance: By carefully designing the distillation process with semantic and perceptual losses, the authors demonstrate that a single-step diffusion model can outperform complex multi-step methods in terms of perceptual quality.

Key Results

  • Quantitative: VPD-SR achieves state-of-the-art (SOTA) performance on both synthetic and real-world datasets, particularly excelling in non-reference perceptual quality metrics like CLIPIQA and MUSIQ. For instance, on the RealSR* dataset, VPD-SR achieves the best CLIPIQA score (0.7635) and the second-best MUSIQ score (63.5923), significantly outperforming the teacher model ResShift.
  • Qualitative: Visual comparisons show that VPD-SR generates images with higher clarity, more natural textures, and better preservation of fine details (e.g., building textures, text, hands, flowers) compared to other methods. Competitors often exhibit excessive sharpening or diminished textural fidelity.
  • Efficiency: VPD-SR achieves this performance with a single inference step, drastically reducing computation time compared to multi-step diffusion models. While its inference time is slightly higher than OSEDiff (which uses LoRA fine-tuning), it offers a more direct distillation approach.

Limitations and Future Work

  • Full-reference metrics: The paper acknowledges that while VPD-SR excels in no-reference metrics, it shows minor limitations in full-reference fidelity metrics.
  • Model size: Although it has fewer parameters than some multi-step diffusion models, VPD-SR still has over 100 million parameters.
  • Future Work: The authors aim to design a more lightweight denoising network and further address the trade-off between fidelity and perceptual quality.

Summary updated

BibTeX
@inproceedings{wu2026one,
  author = {Xue Wu and Jingwei Xin and Jun Hao and Hui Gao and Jie Li and Nannan Wang and Xinbo Gao},
  title = {{One-step diffusion-based real-world image super-resolution with visual perception distillation}},
  booktitle = {Neurocomputing},
  year = {2026},
  doi = {10.1016/j.neucom.2026.133066},
}
Figure