IEEE Transactions on Multimedia (TMM), Accepted on Dec. 30, 2025

• Datasets
  Public benchmark datasets will be organized and released.

• Pretrained Models
  Model weights for all experiments in the paper will be made available.

• Training & Evaluation Code
  Complete training and testing pipelines will be released for reproducibility.

Blind Image Restoration (BIR) aims to recover high-quality (HQ) images from severely degraded low-quality (LQ) inputs with unknown degradations, such as blur, noise, compression artifacts, and low resolution.
Recent BIR methods increasingly rely on vector-quantized (VQ) generative models (e.g., VQGAN) to leverage powerful discrete priors for hallucinating realistic textures under heavy degradation.

Despite their success, existing VQ-based BIR methods inherit a fundamental mismatch between the generation-oriented design of vector quantization and the reconstruction-oriented objective of image restoration. This work revisits the role of vector quantization in BIR and proposes a simple yet effective alternative.

Vector quantization was originally designed to improve image generation quality by mapping continuous features onto a discrete codebook, which helps reduce uncertainty and stabilize generative modeling.
However, blind image restoration prioritizes reconstruction fidelity, requiring accurate and continuous estimation of latent features corresponding to the original HQ image.

This discrepancy leads to a key limitation in existing VQ-based BIR methods:

• The discrete one-hot code selection in VQ introduces quantization errors and information bottlenecks.
• While beneficial for generative diversity, this discreteness restricts the ability to faithfully reconstruct fine-grained image details.
• As a result, restoration performance is often suboptimal, especially under severe degradations.

Through systematic analysis of four representative VQ-based BIR methods, we identify the following observations:

• Observation 1:
  HQ features extracted directly from the HQ encoder reconstruct images significantly better than their VQ-quantized counterparts.

• Observation 2:
  Discrete VQ operations constrain the representation capacity of the codebook and limit reconstruction accuracy in BIR tasks.

These observations indicate that VQ-quantized features are not the optimal learning target for reconstruction-oriented restoration.

To resolve the inconsistency between vector quantization and blind image restoration, we propose Vector Regression (VR) as a continuous alternative to the VQ operation.

Instead of selecting a single codebook entry via hard or soft assignment, VR learns real-valued regression weights over the entire HQ codebook, enabling flexible linear combinations of code vectors to approximate HQ features more accurately.

Key characteristics of VR include:

• Continuous representation:
  Transitions from discrete code selection to continuous feature estimation.
• Regression over codebook:
  Learns weighted combinations of codebook entries rather than one-hot assignments.
• HQ feature supervision:
  The regression module is directly supervised by HQ features extracted during the prior learning stage.

From hard VQ to vector regression: VR enables more flexible and accurate feature approximation.

The main contributions of this work are summarized as follows:

• We reveal the inconsistency between generation-oriented vector quantization and reconstruction-oriented blind image restoration, which has been largely overlooked in prior work.
• We demonstrate that HQ features provide a stronger supervision signal than VQ-quantized features for improving restoration fidelity.
• We propose a simple yet effective Vector Regression (VR) module that replaces VQ operations and consistently improves multiple VQ-based BIR methods across different restoration tasks.

Extensive experiments show that replacing VQ with VR consistently improves:

• Blind face restoration (e.g., CodeFormer, DAEFR)
• Blind image super-resolution (e.g., FeMaSR, AdaCode)

VR achieves higher PSNR and SSIM while maintaining favorable computational efficiency.

The VR module is lightweight, plug-and-play, and can be easily integrated into existing VQ-based blind image restoration frameworks.

If you find this work useful in your research, please consider citing:

@article{Zheng2025VR4BIR,
  title   = {Rethinking the Role of Vector Quantization for Blind Image Restoration},
  author  = {Zheng, Zhaolin and Xue, Liqi and Li, Linghao and Gong, Chenwei and Zhen, Xiantong and Xu, Jun},
  journal = {IEEE Transactions on Multimedia},
  year    = {2025},
  note    = {Accepted}
}