Text to Image Synthesis

Project description: To assess the importance of text embedding in the task of text to image synthesis with GANs, we tested Skip Thoughts embedding and Visual Semantic Embedding as text encodings. Through quantitative and qualitative evaluation, we show that while Skip-Thought seems to impact the performance of a GAN in generating realistic images negatively, a GAN trained with Visual Semantic Embeddings achieves as high performances as the baseline model, strengthening our hypothesis that high performance can be achieved with a simple embedding such as Visual Semantic Embedding as an alternative to baseline model’s embedding which was specifically trained for the task of zero-shot image retrieval.

For more details see report.