The paper introduces the Deep Recurrent Attentive Writer (DRAW) neural network architecture for image generation. DRAW combines a spatial attention mechanism that mimics human eye foveation with a sequential variational auto-encoding framework, allowing for iterative construction of complex images. The system significantly improves upon state-of-the-art generative models on the MNIST dataset and can generate images indistinguishable from real data on the Street View House Numbers dataset. The core of DRAW is an encoder-decoder pair of recurrent neural networks, where the encoder compresses input images and the decoder reconstructs them. The system is trained end-to-end using stochastic gradient descent with a variational upper bound on the log-likelihood of the data. The attention mechanism in DRAW allows the network to selectively focus on parts of the scene while ignoring others, enabling iterative refinement of images. The paper also discusses the read and write operations in DRAW, including a selective attention model inspired by handwriting synthesis and Neural Turing Machines. Experimental results on various datasets demonstrate the effectiveness of DRAW in generating realistic images and improving classification performance.The paper introduces the Deep Recurrent Attentive Writer (DRAW) neural network architecture for image generation. DRAW combines a spatial attention mechanism that mimics human eye foveation with a sequential variational auto-encoding framework, allowing for iterative construction of complex images. The system significantly improves upon state-of-the-art generative models on the MNIST dataset and can generate images indistinguishable from real data on the Street View House Numbers dataset. The core of DRAW is an encoder-decoder pair of recurrent neural networks, where the encoder compresses input images and the decoder reconstructs them. The system is trained end-to-end using stochastic gradient descent with a variational upper bound on the log-likelihood of the data. The attention mechanism in DRAW allows the network to selectively focus on parts of the scene while ignoring others, enabling iterative refinement of images. The paper also discusses the read and write operations in DRAW, including a selective attention model inspired by handwriting synthesis and Neural Turing Machines. Experimental results on various datasets demonstrate the effectiveness of DRAW in generating realistic images and improving classification performance.