PassionSR: Post-Training Quantization
with Adaptive Scale in One-Step
Diffusion based Image Super-Resolution

LiBo Zhu1, Jianze Li1, Haotong Qin2†, Wenbo Li3, Yulun Zhang1†, Yong Guo4, Xiaokang Yang1
1Shanghai Jiao Tong University, 2ETH Zurich,
3Max Planck Institute for Informatics, 4Chinese University of Hong Kong Co., Ltd
†Indicates Corresponding Authors.
Paper Supplementary Code arXiv
Visual Results Pretrained Models

Super Restoration (×4)

Abstract

Diffusion-based image super-resolution (SR) models have shown superior performance at the cost of multiple denoising steps. However, even though the denoising step has been reduced to one, they require high computational costs and storage requirements, making it difficult for deployment on hardware devices. To address these issues, we propose a novel post-training quantization approach with adaptive scale in one-step diffusion (OSD) image SR, PassionSR. First, we simplify OSD model to two core components, UNet and Variational Autoencoder (VAE) by removing the CLIPEncoder. Secondly, we propose Learnable Boundary Quantizer (LBQ) and Learnable Equivalent Transformation (LET) to optimize the quantization process and manipulate activation distributions for better quantization. Finally, we design a Distributed Quantization Calibration (DQC) strategy that stabilizes the training of quantized parameters for rapid convergence. Comprehensive experiments demonstrate that PassionSR with 8-bit and 6-bit obtains comparable visual results with full-precision model. Moreover, our PassionSR achieves significant advantages over recent leading low-bit quantization methods for image SR.

Method

Overview of PassionSR
Step 1: We simplify OSEDiff by removing the DAPE and CLIP encoder modules, resulting in a streamlined full-precision model named PassionSR-FP. Step 2: Our quantizer includes two key trainable components: the Learnable Boundary Quantizer (LBQ) and the Learnable Equivalent Transformation (LET), which enhance the expressiveness and adaptability of the quantization process. Step 3: A distributed calibration strategy and a specially designed loss function are introduced to accelerate convergence and stabilize the quantization optimization.

Results

Quantitative Comparisons (click to expand)
  • Results of UNet and Vae quantization in Table 2 from main paper.

Overview of PassionSR

  • Results of UNet-only quantization in Table 1 from supplementary material.

Overview of PassionSR

Visual Comparisons (click to expand)
  • Results of UNet and Vae quantization in Figure 6 from main paper.

Overview of PassionSR

  • Results of UNet-only quantization in Figure 3 from supplementary material.

Overview of PassionSR

  • More results of UNet and Vae quantization in Figure 1 from supplementary material.

Overview of PassionSR

Poster

PowerPoint

BibTeX

          @inproceedings{zhu2025passionsr,
            title={{PassionSR}: Post-Training Quantization with Adaptive Scale in One-Step Diffusion based Image Super-Resolution},
            author={Zhu, Libo and Li, Jianze and Qin, Haotong and Zhang, Yulun and Guo, Yong and Yang, Xiaokang},
            booktitle={CVPR},
            year={2025}
          }