VOSR: Vision-Only Super-Resolution Framework

When using multi-step models, you can adjust the trade-off between faithfulness to the Low-Resolution (LR) input and generative detail using these parameters:

• --cfg_scale: Controls fidelity to the condition. Higher values emphasize fidelity; lower values allow more generative freedom. A practical range is -0.5 to 2, with 0.5 being a common default.
• --weak_cond_strength_aelq: Controls how much the model follows the condition. Smaller values are more generative; larger values are more faithful. Default is 0.1.
• --infer_steps: Number of sampling steps (default 25).

Benchmark Presets:

• For RealSR: Use --cfg_scale -0.5.
• For ScreenSR: Use --cfg_scale 0.5.

To prevent GPU memory exhaustion when processing large images, use tiling in the latent space. Overlapping tiles are blended using a Gaussian method.

• --tile_size <pixels>: Sets the DiT tile size (default 0 = disabled).
• --vae_tile_size <pixels>: Sets the VAE encoder/decoder tile size (default 0 = full-image). Note that the Qwen VAE is memory-efficient and can handle up to ~8K inputs without tiling.
• --vae_tile_overlap: Sets the VAE tile overlap (defaults to the value of --tile_size).

Example (1.4B one-step with tiling):

python inference_vosr_onestep.py \
    -c preset/ckpts/VOSR_1.4B_os \
    -i large_input.png \
    -o preset/results \
    -u 4 --tile_size 512 --vae_tile_size 1024

To set up the VOSR environment, clone the repository and create a Conda environment with Python 3.8 or higher. Then, install the required dependencies using pip.

## clone this repository
git clone https://github.com/cswry/VOSR.git
cd VOSR

# create an environment with python >= 3.8
conda create -n vosr python=3.8
conda activate vosr
pip install -r requirements.txt

VOSR supports two data loading modes via the dataset_type key in your YAML configuration:

1. txt mode: Each folder contains individual image files (.png / .jpg). You must provide a text configuration file that lists folders with sampling weights.
2. webdataset mode: Each folder contains .tar shards. This uses the same text configuration format as txt mode but loads via WebDataset.

Dataset Config Format: Create a file (e.g., configs/train_txt/train_dataset_txt.txt) where each line follows the pattern: <folder_path>, <sampling_weight>

Example:

/path/to/dataset_A, 2
/path/to/dataset_B, 1
/path/to/dataset_C, 1

Higher weights result in more frequent sampling.

If you use VOSR in your research, please cite the following BibTeX entry:

@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}
}