This paper introduces the Image Transformer, a model based on self-attention mechanisms for image generation. The model is designed to generate images with a tractable likelihood, enabling effective training and evaluation. By restricting self-attention to local neighborhoods, the model can process larger images while maintaining a larger receptive field than typical convolutional neural networks. The Image Transformer significantly outperforms the current state of the art in image generation on ImageNet, improving the best published negative log-likelihood from 3.83 to 3.77. The model is also applied to image super-resolution with a large magnification ratio, achieving better results than previous methods. In a human evaluation study, images generated by the super-resolution model fool human observers three times more often than the previous state of the art. The Image Transformer is based on a self-attention mechanism, which allows for parallel processing and efficient training. The model uses a local self-attention mechanism to restrict the attention to local neighborhoods, enabling the model to process larger images. The model is evaluated on various tasks, including image generation, image super-resolution, and conditional image generation. The results show that the Image Transformer achieves significant improvements in image generation quality and performance. The model is trained using maximum likelihood, with the network outputting all parameters of the autoregressive distribution. The model uses two different settings of the distribution: a categorical distribution across each channel and a mixture of discretized logistics over three channels. The model is evaluated on the CIFAR-10 and ImageNet datasets, achieving state-of-the-art results in image generation. The model is also applied to image super-resolution, achieving better results than previous methods. The results show that the Image Transformer is effective in generating realistic images and improving the quality of image super-resolution. The model is also evaluated on conditional image generation, where it is conditioned on image classes and super-resolution inputs. The results show that the model can generate realistic images for various categories and super-resolution inputs. The model is also evaluated on human evaluation studies, where it is found to be more effective in generating realistic images than previous methods. The model is also evaluated on the consistency of high-resolution samples with low-resolution inputs, showing that the model generates high-quality images by adding synthesized details on the low-resolution input image. The model is also evaluated on the performance of different attention mechanisms, showing that 2D local attention with a sampling temperature of 0.9 achieves the highest perceptual quality in human evaluation studies. The model is also evaluated on the performance of different architectures, showing that the Image Transformer is effective in generating realistic images and improving the quality of image super-resolution. The model is also evaluated on the performance of different training settings, showing that the Image Transformer is effective in generating realistic images and improving the quality of image super-resolution. The model is also evaluated on the performance of different data sets, showing that the Image Transformer is effective in generating realistic images and improving the quality of image super-resolution.This paper introduces the Image Transformer, a model based on self-attention mechanisms for image generation. The model is designed to generate images with a tractable likelihood, enabling effective training and evaluation. By restricting self-attention to local neighborhoods, the model can process larger images while maintaining a larger receptive field than typical convolutional neural networks. The Image Transformer significantly outperforms the current state of the art in image generation on ImageNet, improving the best published negative log-likelihood from 3.83 to 3.77. The model is also applied to image super-resolution with a large magnification ratio, achieving better results than previous methods. In a human evaluation study, images generated by the super-resolution model fool human observers three times more often than the previous state of the art. The Image Transformer is based on a self-attention mechanism, which allows for parallel processing and efficient training. The model uses a local self-attention mechanism to restrict the attention to local neighborhoods, enabling the model to process larger images. The model is evaluated on various tasks, including image generation, image super-resolution, and conditional image generation. The results show that the Image Transformer achieves significant improvements in image generation quality and performance. The model is trained using maximum likelihood, with the network outputting all parameters of the autoregressive distribution. The model uses two different settings of the distribution: a categorical distribution across each channel and a mixture of discretized logistics over three channels. The model is evaluated on the CIFAR-10 and ImageNet datasets, achieving state-of-the-art results in image generation. The model is also applied to image super-resolution, achieving better results than previous methods. The results show that the Image Transformer is effective in generating realistic images and improving the quality of image super-resolution. The model is also evaluated on conditional image generation, where it is conditioned on image classes and super-resolution inputs. The results show that the model can generate realistic images for various categories and super-resolution inputs. The model is also evaluated on human evaluation studies, where it is found to be more effective in generating realistic images than previous methods. The model is also evaluated on the consistency of high-resolution samples with low-resolution inputs, showing that the model generates high-quality images by adding synthesized details on the low-resolution input image. The model is also evaluated on the performance of different attention mechanisms, showing that 2D local attention with a sampling temperature of 0.9 achieves the highest perceptual quality in human evaluation studies. The model is also evaluated on the performance of different architectures, showing that the Image Transformer is effective in generating realistic images and improving the quality of image super-resolution. The model is also evaluated on the performance of different training settings, showing that the Image Transformer is effective in generating realistic images and improving the quality of image super-resolution. The model is also evaluated on the performance of different data sets, showing that the Image Transformer is effective in generating realistic images and improving the quality of image super-resolution.