Principle:Junyanz Pytorch CycleGAN and pix2pix Pretrained Model Download
| Knowledge Sources | pytorch-CycleGAN-and-pix2pix |
|---|---|
| Domains | Image-to-Image Translation, Model Distribution, Transfer Learning |
| Last Updated | 2026-02-09 |
Overview
A model acquisition step that downloads pre-trained generator checkpoints for immediate image translation inference without training.
Description
The repository provides shell scripts to download pre-trained model weights hosted on Berkeley servers. These weights allow users to perform inference immediately without needing to train models from scratch, which can take hours or days on a GPU.
Downloaded checkpoints are saved as latest_net_G.pth files within ./checkpoints/<model_name>_pretrained/ directories, following the standard checkpoint naming convention used by the model loading system.
Available CycleGAN pretrained models:
- horse2zebra -- Translates horses to zebras and vice versa
- monet2photo -- Translates between Monet paintings and photographs
- style_monet -- Applies Monet painting style
- style_cezanne -- Applies Cezanne painting style
- style_ukiyoe -- Applies Ukiyo-e art style
- style_vangogh -- Applies Van Gogh painting style
- apple2orange -- Translates between apples and oranges
- summer2winter_yosemite -- Translates between summer and winter Yosemite scenes
Available pix2pix pretrained models:
- edges2shoes -- Generates shoes from edge drawings
- sat2map -- Converts satellite imagery to maps
- map2sat -- Converts maps to satellite imagery
- facades_label2photo -- Generates building facades from label maps
- night2day -- Translates night scenes to daytime
Usage
Run the appropriate download script before running inference with test.py. The downloaded model name must match the --name argument passed to the test script, with _pretrained appended.
Theoretical Basis
Pre-trained model distribution is a standard practice in deep learning that democratizes access to trained models. For GAN models in particular, training requires careful hyperparameter tuning, substantial compute resources, and significant time. Providing pre-trained checkpoints allows:
- Immediate inference for users who want to apply image translation without training
- Reproducibility by providing the exact model weights used to generate published results
- Transfer learning as a starting point for fine-tuning on related domains
- Evaluation baselines for comparing new methods against established results