text2art-gan

Project aimed to create a GAN, that is able to generate art just from text descriptions. The project can be accessed from GitLab or GitHub.

CUB

For each style, the first column is where lambda_idt = 0.5 and the second column is where lambda_idt = 5.

COCO

For each style, the first column is where lambda_idt = 0.5 and the second column is where lambda_idt = 5.

When given a more abstract word, the results can be quite interesting:

Set Up

Dependencies

• python 3.7
• pytorch
• numpy
• easydict
• pyyaml

Model parameters

You can download the model parameters from here.

Installation

• Install dependecies
• Clone this repo:

git clone https://gitlab.com/karl-joan/text2art-gan.git
cd text2art-gan

• Download model parameters to the current working directory
• Extract model parameters

7z x parameters.7z

Usage

$ python generate.py --help
usage: generate.py [-h] [-d {birds,coco}] [-n NUMBER] [-i] [-c] [-v]
                   "caption" {abstract_expressionism,impressionism}

Generate art from text

positional arguments:
  "caption"             text to generate from
  {abstract_expressionism,impressionism}
                        the style of the artwork

optional arguments:
  -h, --help            show this help message and exit
  -d {birds,coco}, --dataset {birds,coco}
                        dataset to generate from (default birds)
  -n NUMBER, --number NUMBER
                        the number of artworks to generate (default 2)
  -i, --identity        set lambda_idt = 5 instead of lambda_idt = 0.5
  -c, --cpu             use cpu
  -v, --verbose         print more details

$ python generate -d coco "a man is standing infront of a building" abstract_expressionism

Citation

If you use the work present in this repository for your project, then please consider citing this work.

Acknowledgments

Modified code is borrowed from AttnGAN and CycleGAN.