Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Preliminary Study on GPT-Image Generation Model for Image Restoration

Hao Yang, Yan Yang, Ruikun Zhang, Liyuan Pan

TL;DR

This work systematically evaluates GPT-Image as a restoration prior across diverse degradations, revealing strong perceptual quality but limited pixel-level fidelity due to geometric distortions and misalignments. It proposes a plug-in baseline that uses GPT-Image outputs as priors, aligned with degraded inputs via a deformable Alignment module and refined by a Restormer-based restoration network, achieving improved perceptual scores while maintaining competitive PSNR/SSIM. The approach generalizes across multiple backbones and degradation tasks, suggesting a practical path to combine large multimodal generative priors with traditional restoration pipelines. The findings highlight a promising direction for bridging image generation models and low-level vision tasks, with trade-offs in computation and fidelity that warrant further research and optimization.

Abstract

Recent advances in OpenAI's GPT-series multimodal generation models have shown remarkable capabilities in producing visually compelling images. In this work, we investigate its potential impact on the image restoration community. We provide, to the best of our knowledge, the first systematic benchmark across diverse restoration scenarios. Our evaluation shows that, while the restoration results generated by GPT-Image models are often perceptually pleasant, they tend to lack pixel-level structural fidelity compared with ground-truth references. Typical deviations include changes in image geometry, object positions or counts, and even modifications in perspective. Beyond empirical observations, we further demonstrate that outputs from GPT-Image models can act as strong visual priors, offering notable performance improvements for existing restoration networks. Using dehazing, deraining, and low-light enhancement as representative case studies, we show that integrating GPT-generated priors significantly boosts restoration quality. This study not only provides practical insights and a baseline framework for incorporating GPT-based generative priors into restoration pipelines, but also highlights new opportunities for bridging image generation models and restoration tasks. To support future research, we will release GPT-restored results.

A Preliminary Study on GPT-Image Generation Model for Image Restoration

TL;DR

This work systematically evaluates GPT-Image as a restoration prior across diverse degradations, revealing strong perceptual quality but limited pixel-level fidelity due to geometric distortions and misalignments. It proposes a plug-in baseline that uses GPT-Image outputs as priors, aligned with degraded inputs via a deformable Alignment module and refined by a Restormer-based restoration network, achieving improved perceptual scores while maintaining competitive PSNR/SSIM. The approach generalizes across multiple backbones and degradation tasks, suggesting a practical path to combine large multimodal generative priors with traditional restoration pipelines. The findings highlight a promising direction for bridging image generation models and low-level vision tasks, with trade-offs in computation and fidelity that warrant further research and optimization.

Abstract

Recent advances in OpenAI's GPT-series multimodal generation models have shown remarkable capabilities in producing visually compelling images. In this work, we investigate its potential impact on the image restoration community. We provide, to the best of our knowledge, the first systematic benchmark across diverse restoration scenarios. Our evaluation shows that, while the restoration results generated by GPT-Image models are often perceptually pleasant, they tend to lack pixel-level structural fidelity compared with ground-truth references. Typical deviations include changes in image geometry, object positions or counts, and even modifications in perspective. Beyond empirical observations, we further demonstrate that outputs from GPT-Image models can act as strong visual priors, offering notable performance improvements for existing restoration networks. Using dehazing, deraining, and low-light enhancement as representative case studies, we show that integrating GPT-generated priors significantly boosts restoration quality. This study not only provides practical insights and a baseline framework for incorporating GPT-based generative priors into restoration pipelines, but also highlights new opportunities for bridging image generation models and restoration tasks. To support future research, we will release GPT-restored results.
Paper Structure (7 sections, 8 figures, 4 tables)

This paper contains 7 sections, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Restoration of Diverse Degradation
  4. Failure Cases
  5. A Baseline Solution
  6. Discussion
  7. Conclusion

Figures (8)

  • Figure 1: Image restoration results of GPT-Image on real world degradation without available ground truth. The first row and second row are degraded inputs and the restored outputs, respectively. (a)-(d) correspond to low-light conditions, heavy noise, motion blur, and dense haze, respectively.
  • Figure 2: Image restoration results of GPT-Image on real-world degraded images without ground truth. Each vertical pair shows a degraded input image (top) and its corresponding restored output (bottom), with the type of degradation labeled beside each pair.
  • Figure 3: Image restoration results of GPT-Image on real-world degraded images with available ground truth. Each triplet consists of the ground truth image, the degraded input, and the corresponding restored output, with the type of degradation labeled beside each set. We display the PSNR and CLIP-IQA scores below each image, reflecting perceptual quality and pixel-level structural fidelity, respectively.
  • Figure 4: Failure cases. (a) Variations in image proportions. (b) Shifts in object positions and quantities. (c) Changes in viewpoint.
  • Figure 5: Pipeline of our proposed solution. (a) Overall Pipeline, (b) the structure of Alignment Module. We use GPT-Image-generated images as priors, align them with the degraded image using a deformable convolution-based alignment module, and feed the aligned features together with the degraded image into a restoration backbone to obtain the final restored result.
  • ...and 3 more figures