VOSR: A Vision-Only Generative Model for Image Super-Resolution
A framework for native generative image restoration.
Many recent generative super-resolution methods adapt pretrained text-to-image models. VOSR trains its generative backbone from scratch for super-resolution, using structural and visual semantic conditions from the low-resolution input. It requires no text prompts and reuses pretrained VAE and vision-encoder components.
Our goal is to build native generative models for image restoration that combine input fidelity, perceptual quality, and efficient inference. VOSR explores this direction through restoration-oriented guidance and both multi-step and one-step models.
Available checkpoints
This repository provides:
VOSR2/: VOSR 2.0, a 1.4B one-step model.VOSR_0.5B_ms/: 0.5B multi-step model.VOSR_0.5B_os/: 0.5B one-step model.VOSR_1.4B_ms/: 1.4B multi-step model.VOSR_1.4B_os/: 1.4B one-step model.
Supporting VAE, decoder, and visual-encoder files are also provided. See the license section for the terms that apply to each component.
Qwen-Image-vae-2d/ is our 2D conversion of the released 3D causal VAE weights, with the encoder from Wan2.1 and the image decoder fine-tuned by Qwen-Image.
Intended use and limitations
VOSR is intended for research and image-restoration applications, including single-image super-resolution.
Generative super-resolution may produce plausible details that are not fully supported by the input image. The output should not be treated as a faithful recovery of information absent from the input or as forensic evidence. Results may vary with the input image, degradation, checkpoint, and inference settings.
Training and evaluation
The accompanying paper describes the training setup and evaluation protocols. It reports training on a filtered image collection of approximately 100 million images and using synthetic low-resolution/high-resolution pairs generated with Real-ESRGAN degradation. The training data is not included in this repository.
See the paper for evaluation results and further details.
Inference
Use the inference code in the VOSR project repository. This repository uses VOSR-specific inference scripts; the generic DiffusionPipeline.from_pretrained(...) example shown by Hugging Face may not apply.
Clone the project and install its dependencies:
git clone https://github.com/cswry/VOSR.git
cd VOSR
pip install -r requirements.txt
Download the required checkpoint files from this Hugging Face repository into the project’s preset/ckpts/ directory, preserving their folder names. For example, to run VOSR 2.0 on images in preset/datasets/inp_data:
python inference_vosr_onestep.py \
-c preset/ckpts/VOSR2 \
-i preset/datasets/inp_data \
-o preset/results \
-u 4
For inference with other checkpoints and options, see the project README.
License
The Apache-2.0 license applies to VOSR project materials covered by the project license, including project-authored checkpoints unless a release-specific notice says otherwise. Some bundled third-party components have their own terms; the Apache-2.0 license does not replace those terms.
| Files | Terms |
|---|---|
VOSR checkpoints (VOSR2/, VOSR_*/) and project-authored decoder files | Apache-2.0 under the project license, unless a release-specific notice says otherwise. |
Qwen-Image-vae-2d/ | Follow the applicable upstream terms of Wan2.1 and Qwen-Image. |
stable-diffusion-2-1-base/ | Follow the upstream Stable Diffusion 2.1 terms, listed as OpenRAIL++. |
torch_cache/ | If this directory contains DINOv2 weights, follow the terms in the DINOv2 repository. |
Please retain applicable upstream notices and comply with the terms for each third-party component.
Citation
If you use VOSR in your research, please cite:
@inproceedings{wu2026vosr,
title = {VOSR: A Vision-Only Generative Model for Image Super-Resolution},
author = {Wu, Rongyuan and Sun, Lingchen and Zhang, Zhengqiang and Kong, Xiangtao and Zhao, Jixin and Wang, Shihao and Zhang, Lei},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year = {2026}
}