English

Fine-structure Preserved Real-world Image Super-resolution via Transfer VAE Training

Computer Vision and Pattern Recognition 2025-07-29 v1

Abstract

Impressive results on real-world image super-resolution (Real-ISR) have been achieved by employing pre-trained stable diffusion (SD) models. However, one critical issue of such methods lies in their poor reconstruction of image fine structures, such as small characters and textures, due to the aggressive resolution reduction of the VAE (eg., 8×\times downsampling) in the SD model. One solution is to employ a VAE with a lower downsampling rate for diffusion; however, adapting its latent features with the pre-trained UNet while mitigating the increased computational cost poses new challenges. To address these issues, we propose a Transfer VAE Training (TVT) strategy to transfer the 8×\times downsampled VAE into a 4×\times one while adapting to the pre-trained UNet. Specifically, we first train a 4×\times decoder based on the output features of the original VAE encoder, then train a 4×\times encoder while keeping the newly trained decoder fixed. Such a TVT strategy aligns the new encoder-decoder pair with the original VAE latent space while enhancing image fine details. Additionally, we introduce a compact VAE and compute-efficient UNet by optimizing their network architectures, reducing the computational cost while capturing high-resolution fine-scale features. Experimental results demonstrate that our TVT method significantly improves fine-structure preservation, which is often compromised by other SD-based methods, while requiring fewer FLOPs than state-of-the-art one-step diffusion models. The official code can be found at https://github.com/Joyies/TVT.

Keywords

Cite

@article{arxiv.2507.20291,
  title  = {Fine-structure Preserved Real-world Image Super-resolution via Transfer VAE Training},
  author = {Qiaosi Yi and Shuai Li and Rongyuan Wu and Lingchen Sun and Yuhui Wu and Lei Zhang},
  journal= {arXiv preprint arXiv:2507.20291},
  year   = {2025}
}

Comments

ICCV 2025

R2 v1 2026-07-01T04:21:00.657Z