Text2Image

Goal : In terms of specific text description to generate corresponding images.
TensorFlow implement for Kaggles Contest - Reverse Image Caption
Technique: Using GAN-CLS algorithm from the paper Generative Adversarial Text-to-Image Synthesis and stackGAN++ model from StackGAN++: Realistic Image Synthesis with Stacked Generative Adversarial Networks

Download Data and Pre-trained Model

Download image files, captions, and pre-trained model from Google Drive. Image and captions file put into ./text-to-image directory. Moreover, the pre-trained model put into ./checkpoint directory.

the flower shown has yellow anther red pistil and bright red petals.
this flower has petals that are yellow, white and purple and has dark lines
the petals on this flower are white with a yellow center
this flower has a lot of small round pink petals.
this flower is orange in color, and has petals that are ruffled and rounded.
the flower has yellow petals and the center of it is brown
this flower has petals that are blue and white.
these white flowers have petals that start off white in color and end in a white towards the tips.

Use stackGAN can get better result, despite the improved-wgan with skip-thought also can produce satisfying one.
Get more captions per images, that is, randomly choose 3~5 captions for each picture (if we choose all captions, the training set is so large that it takes long time to train).
Add wrong image with false label to train with Discriminator: That is, give a caption and a random image from dataset, and give it false label. (To let D learn whether the image has relation with the caption.)
The result of random distortion(e.g. brightness and different angles) to images is not good as I think.