Paper summary

Semantic Image Synthesis via Adversarial Learning

1.What is this paper about?

It proposes an end-to-end neural architecture(SIS-GAN) that leverages adversarial learning to automatically learn implicit loss functions, to manipulate the image from text description while keeping the image realistic.

2.What’s better than previous paper?

Previous model aim to synthesize realistic images directly from natural language descriptions. It is first model to manipulate a source image semantically with text descriptions, while still maintain features that are irrelevant to what text descriptions.

3.What are important parts of technique and methods?

???

4.How did they verify it?

It is evaluated by cunducting two experiments on Caltech-200 bird dataset and Oxford-102 flower dataset. It is first model to manipulate the images from text description, so it is hard to compare, but to compare to baseline method it shows the effectiveness.

Compared with the baseline model, its result kept most of the original background, pose, shape and other information in the original images while keeping the image realistic.

It use human evaluation to compare the baseline model.(1(best) to 3(worst))

• Whetherthesynthesizedimagekeepstheoriginalpose of the bird / flower; • Whether the synthesized image keeps the original background; • Whether the synthesized image matches the text de- scription while being realistic.

Its result shows it outperfome baseline model and its model without VGG in all of metrics.