Image
pytorch-CycleGAN-and-pix2pix
The classic PyTorch implementation of CycleGAN and pix2pix for image-to-image translation
★ 25.3k starsNOASSERTION
View guide →CUT — contrastive unpaired image-to-image translation in PyTorch, faster and lighter than CycleGAN (ECCV 2020)
contrastive-unpaired-translation is an open-source image project by taesungp (taesungp/contrastive-unpaired-translation on GitHub). It has 2,556 GitHub stars and 452 forks, is released under the NOASSERTION license, and was last updated on GitHub on 5 September 2023.
Gather two unpaired image collections (domains A and B) — e.g. horses and zebras — or download one of the provided datasets; decide CUT vs FastCUT vs single-image SinCUT.
Clone into a folder named CUT, as the README's getting-started shows.
pip install -r requirements.txt (or conda env create -f environment.yml); optionally start a visdom server to watch training.
Download a dataset with the provided script, train with train.py, then generate results with test.py — or skip training using the pretrained models and launcher scripts.
Patchwise contrastive learning ties corresponding patches of input and output together, replacing cycle consistency and its inverse network.
The authors report faster training and lower memory than CycleGAN, with stronger distribution matching in their experiments.
A lighter configuration (no identity loss, higher lambda_NCE) described as using half the GPU memory and training twice as fast as CycleGAN.
Each 'domain' can be just one image, using StyleGAN2-based components adopted in the repo.
Downloadable pretrained checkpoints (Cityscapes, Horse2Zebra, Cat2Dog) with launcher scripts and FID evaluation instructions.
Experiment launchers generate the long train/test command lines for the provided datasets.
Clone CUT into a directory named CUT, as the README shows.
git clone https://github.com/taesungp/contrastive-unpaired-translation.git
cd contrastive-unpaired-translationInstall PyTorch and the other requirements with pip (or use the conda environment file).
pip install -r requirements.txtDownload the example dataset, train CUT, then run the test script.
bash ./datasets/download_cut_dataset.sh grumpifycat
python train.py --dataroot ./datasets/grumpifycat --name grumpycat_CUT --CUT_mode CUT
python test.py --dataroot ./datasets/grumpifycat --name grumpycat_CUT --CUT_mode CUT --phase trainThe authors — Taesung Park, Alexei A. Efros, Richard Zhang and Jun-Yan Zhu (UC Berkeley / Adobe Research) — published it with their ECCV 2020 paper.
The code is public research code; GitHub lists its license as NOASSERTION, so read the repository's license file before reusing it commercially.
No — training and inference run locally in PyTorch on your own GPU.
CycleGAN enforces cycle consistency with a second generator. CUT uses a contrastive PatchNCE loss instead, which the authors show trains faster with less memory; FastCUT is the even lighter option.
Clone, pip install -r requirements.txt, download the grumpifycat dataset, and run the train.py command from the README.
The classic PyTorch implementation of CycleGAN and pix2pix for image-to-image translation
🤗 Diffusers: State-of-the-art diffusion models for image, video, and audio generation in PyTorch.
Unofficial DragGAN implementation: drag points on a GAN image to edit it interactively
Local-first AI e-commerce image suite generator for solo sellers
Modern AI image lab on Pollinations.ai — models, seeds, history, transparency
Photoshop-ai-smart-enhance — ai-powered image enhancement extension for adobe photos, o...
The most-starred other Image projects in the LibreAI catalogue — the realistic shortlist if contrastive-unpaired-translation is not the right fit.
Stable Diffusion web UI
The node-graph powerhouse for precise, professional AI image and video generation.
High performance self-hosted photo and video management solution
Focus on prompting and generating
🆙 Upscayl - #1 Free and Open Source AI Image Upscaler for Linux, MacOS and Windows
Free and open — grab the source, star the repo, and start creating.