Semi-supervised FusedGAN for Conditional Image Generation

1.What is this paper about?

It proposes FusedGAN, a deep network for conditional image synthesis with controllable sampling of diverse images.

2.What’s better than previous paper?

Previouse one are insufficient in Fidelity, diversity and controllable aspects. It can perform controllable sampling of diverse images with very high fidelity by disentangling the generation process into various stages.

3.What are important parts of technique and methods?

we propose to disentangle the image generation process such that we learn two cascaded generators. The first unconditional generator produces a structure prior which is independent of the condition, and the second conditional generator further adds style to it and creates an image that matches the condition.

4.How did they verify it?

It analysis two conditional image generation use cases.

• Text-to-image synthesis using bird dataset

It valid on CUB birds dataset and nabirds dataset(because it is semi-supervised). For quantitative results, we use the inception score.

In both the case of fixed posture with varying styles and fixed posture with varying details, it is able to generate images that met the requirements. In addition, high performance was obtained in evaluation by metrics and also by human evaluation.

• Attributes-to-image synthesis using face dataset

It use celebA dataset for analyses. By sampling with the same structure but varying attributes, it shows to be able to generate various image which conditioned by text description.