This paper investigates a problem of generating images from visual attributes. Given the prevalent research for image recognition, the conditional image generation problem is relatively under-explored due to the challenges of learning a good generative model and handling rendering uncertainties in images. To address this, we propose a variety of attribute-conditioned deep variational auto-encoders that enjoy both effective representation learning and Bayesian modeling, from which images can be generated from specified attributes and sampled latent factors. We experiment with natural face images and demonstrate that the proposed models are capable of generating realistic faces with diverse appearance. We further evaluate the proposed models by performing attribute-conditioned image progression, transfer and retrieval. In particular, our generation method achieves superior performance in the retrieval experiment against traditional nearest-neighbor-based methods both qualitatively and quantitatively.

探究了从视觉属性生成图像的新问题，开发了一种具有分离潜在变量的分层生成模型，该模型可以使用变分自动编码器进行端到端学习，使用一般的能量最小化算法进行后验推理，展示了在面部和鸟类自然图像中生成具有分离潜在表示的逼真多样样本的优异结果。

Attribute2Image: 从视觉特征生成有条件的图像