Awesome

Semantic Image Synthesis via Adversarial Learning

This is a PyTorch implementation of the paper MC-GAN: Multi-conditional Generative Adversarial Network for Image (BMVC 2018).

Model architecture

Requirements

python 2.7
PyTorch 0.3
Torchvision
easydict
pandas
torchtext
NLTK

Model1

please download a pre-trained model bird and base image for training base_img_bird, base_img_flower

Additional Datasets

Prepare the Caltech-200 birds dataset and Oxford-102 flowers dataset for reproducing main results in this repository

Run Model1

Train a MC-GAN model on the bird (CUB) dataset using our preprocessed embeddings: python main.py --cfg cfg/birds_3stages.yml --gpu 0

Test a MC-GAN model

Change cfg/*.yml files to generate images from pre-trained models.

Train.flag = False
Train.net_G = 'path of pre-trianed model'

result

Brids Brids Brids Brids

Model2

please download a pre-trained model bird and text embedding model

Run Model2

Text embdding
- run train_text_embeddng.py
Train Generative model
- run train_MCb.py
Test
- run test.py

please follow result in this repository

Results

Birds Birds Birds Birds Birds Birds Birds Birds

##Citing MC-GAN

If you find MC-GAN useful in your research, please consider citing:

@inproceedings{park2018mc,
  title={MC-GAN: Multi-conditional Generative Adversarial Network for Image Synthesis},
  author={Park, Hyojin and Yoo, Youngjoon and Kwak, Nojun},
  booktitle={The British MachineVision Conference (BMVC)},
  year={2018}
}