Practical overview
When it fits
Use it when you need a locally run, script-based motion-deblurring model or want to explore training with paired blurry and sharp images.
Who it is for
- Developers and researchers working with image restoration and computer vision.
- Users with a local PyTorch environment and an NVIDIA GPU who want to run inference or train on paired image data.
Documented uses
- Generate sharp estimates from blurry images using a pretrained model.
- Prepare paired images from two folders and start training on custom data.
- Experiment with the model architecture for related image-to-image tasks, including super-resolution, colorization, inpainting, and dehazing.
Requirements & limits
The implementation is publicly available on GitHub. Running the documented workflow requires a local setup; the README specifies an NVIDIA GPU with CUDA and cuDNN and PyTorch, and says CPU use is untested.
- Inference requires downloaded weights; the README says to retain only generator weights for that stage.
- The documented training path expects paired images and includes a script to combine two image folders.
- The README describes related image-to-image applications as possible uses, but gives no separate workflows for them.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run inference on a folder of blurry images
Apply the documented pretrained generator to images in a folder.
Before you start
- Python and PyTorch
- NVIDIA GPU with CUDA and cuDNN; CPU use is untested in the README
- Downloaded model weights, with only generator weights retained for inference
- Blurry input images in a folder
Steps
Download the model weights linked from the README.
Place the generator weights in the documented
/.checkpoints/experiment_namedirectory.Put the blurry images in a folder and run the command below, substituting the path to the image folder for the documented placeholder.
python test.py --dataroot /.path_to_your_data --model test --dataset_mode single --learn_residual
Expected outcome
The model produces sharp estimates from the blurry images provided as input.
Guide limitations
- The README reports CPU use as untested.
- The command and instructions document local inference, not a hosted batch service or user interface.
Prepare paired images and start model training
Create paired training data from two image folders and launch the documented training script.
Before you start
- Python and PyTorch
- NVIDIA GPU with CUDA and cuDNN; CPU use is untested in the README
- Corresponding images arranged in separate folders for the two sides of each pair
Steps
Run the command below, substituting the relevant folder paths.
python datasets/combine_A_and_B.py --fold_A /path/to/data/A --fold_B /path/to/data/B --fold_AB /path/to/dataRun the command below, substituting the training-data path and a crop size.
python train.py --dataroot /.path_to_your_data --learn_residual --resize_or_crop crop --fineSize CROP_SIZE
Expected outcome
The pairing script prepares image pairs, and the training script starts training on the prepared data.
Guide limitations
- The README does not specify a guaranteed training outcome or resulting model quality; it notes that results depend on the data and training setup.
- CPU use is untested in the README.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- KupynOrest/DeblurGAN · Official homepage title · retained in its source language