Faithful Extreme Image Rescaling with Learnable Reversible Transformation and Semantic Priors
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914524382101504 |
|---|---|
| author | Wei, Hao Zhou, Yanhui Ge, Chenyang Anwar, Saeed Mian, Ajmal |
| author_facet | Wei, Hao Zhou, Yanhui Ge, Chenyang Anwar, Saeed Mian, Ajmal |
| contents | Most recent extreme rescaling methods struggle to preserve semantically consistent structures and produce realistic details, due to the severely ill-posed nature of low- to high-resolution mapping under scaling factors of $16\times$ or higher. To alleviate the above problems, we propose FaithEIR, a diffusion-based framework for extreme image rescaling. Inspired by singular value decomposition, we develop learnable reversible transformation that enables invertible downscaling and upscaling in the latent space. To compensate for information loss due to quantization, we propose an adaptive detail prior, a high-frequency dictionary that captures the empirical average of commonly occurring structures in the training data. Finally, we design a lightweight pixel semantic embedder to provide semantic conditioning for the pretrained diffusion model. We present extensive experimental results demonstrating that our FaithEIR consistently outperforms state-of-the-art methods, achieving superior reconstruction fidelity and perceptual quality. Our code, model weights, and detailed results are released at https://github.com/cshw2021/FaithEIR. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_00605 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Faithful Extreme Image Rescaling with Learnable Reversible Transformation and Semantic Priors Wei, Hao Zhou, Yanhui Ge, Chenyang Anwar, Saeed Mian, Ajmal Computer Vision and Pattern Recognition Most recent extreme rescaling methods struggle to preserve semantically consistent structures and produce realistic details, due to the severely ill-posed nature of low- to high-resolution mapping under scaling factors of $16\times$ or higher. To alleviate the above problems, we propose FaithEIR, a diffusion-based framework for extreme image rescaling. Inspired by singular value decomposition, we develop learnable reversible transformation that enables invertible downscaling and upscaling in the latent space. To compensate for information loss due to quantization, we propose an adaptive detail prior, a high-frequency dictionary that captures the empirical average of commonly occurring structures in the training data. Finally, we design a lightweight pixel semantic embedder to provide semantic conditioning for the pretrained diffusion model. We present extensive experimental results demonstrating that our FaithEIR consistently outperforms state-of-the-art methods, achieving superior reconstruction fidelity and perceptual quality. Our code, model weights, and detailed results are released at https://github.com/cshw2021/FaithEIR. |
| title | Faithful Extreme Image Rescaling with Learnable Reversible Transformation and Semantic Priors |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2605.00605 |