We introduce SinGAN, an unconditional generative model that can be learned from a single natural image. Our model is trained to capture the internal distribution of patches within the image, and is then able to generate high quality, diverse samples that carry the visual content of the image, yet allow for new object configurations and structures to emerge. SinGAN contains a pyramid of fully convolutional GANs, each responsible for learning the patch distribution at a different scale of the image. This allows generating new samples of arbitrary size and aspect ratio, which have significant variability, yet maintain both the global structure and the fine textures of the training image. Compared to previous single image GAN schemes, our approach is not limited to texture images, and is applicable to general natural images.

Generative Adversarial Networks (GANs) have shown remarkable success in learning complex image distributions from large training sets. However, these models typically require thousands to millions of training samples drawn from the same domain. In many practical scenarios, only a single example image is available. The goal of this work is to learn an unconditional generative model from such a single image — capturing its internal patch statistics and generating new, diverse samples that resemble the original image in terms of visual content.

The power of internal image statistics has been long recognized in classical computer vision — patch recurrence across scales within a single image has been exploited for tasks like super-resolution, denoising, and retargeting. Our model builds on this observation: SinGAN learns a multi-scale, unconditional generative model by capturing the internal patch distribution at multiple scales. Each generator in the pyramid is responsible for learning the patch distribution at a different scale, with "each responsible for learning the patch distribution at a different scale." This framework is purely generative (i.e., maps noise to image samples), unlike prior conditional single-image methods that required input images.

Single Image Deep Models. Previous approaches to learning from single images have been task-specific — networks overfitted to a single image for super-resolution (ZSSR), segmentation, or dehazing. In the generative domain, texture synthesis methods learn from single textures but are limited to homogeneous texture patterns. InGAN can retarget a single image but is a conditional GAN requiring an input image, whereas "our framework is purely generative (i.e., maps noise to image samples)." On the other hand, GAN-based image manipulation methods, while powerful, typically require class-specific training datasets, limiting their applicability to specific domains.

SinGAN consists of a pyramid of generators {G_0, ..., G_N}, trained against a pyramid of downsampled versions of the training image {x_0, ..., x_N}, where x_0 is the coarsest scale. "The generation of an image sample starts at the coarsest scale and sequentially passes through all generators." At each scale n, the generator G_n takes as input a noise map z_n and the upsampled output from the previous scale. The architecture of each generator is a fully convolutional network with a Markovian (PatchGAN) discriminator D_n that classifies overlapping patches as real or generated.

Training is done sequentially from the coarsest to the finest scale. Once G_n is trained, it is fixed and G_{n+1} is trained. The loss at each scale combines an adversarial loss and a reconstruction loss: "min_{G_n} max_{D_n} L_adv(G_n, D_n) + alpha * L_rec(G_n)." The adversarial loss employs WGAN-GP, where the discriminator operates on overlapping patches with final scores averaged across the patch discrimination map. The reconstruction loss ensures that a specific set of noise maps can regenerate the original training image, which is essential for enabling image manipulation applications such as super-resolution, harmonization, and editing.

Quantitative evaluation is performed using Amazon Mechanical Turk (AMT) user studies and a novel Single Image Frechet Inception Distance (SIFID) metric. AMT studies show a confusion rate of "21.45% +/- 1.5%" for full generation and "30.45% +/- 1.5%" when starting from coarser scales, indicating that generated samples are often indistinguishable from real patches. The SIFID metric, which measures internal patch statistics preservation rather than dataset-level distributions, shows strong correlation with human perception studies.

SinGAN demonstrates broad applicability across multiple image manipulation tasks. For super-resolution, the model achieves visual quality that exceeds state-of-the-art internal methods, matching SRGAN performance while training on a single image. Paint-to-image conversion transforms clipart into photorealistic images that preserve global structure. Harmonization realistically blends pasted objects by matching background patch distributions. Editing enables seamless compositing through patch distribution matching, producing superior results compared to content-aware tools. Single image animation creates video from a single image through noise-space traversal.

SinGAN enables unconditional generation from single natural images, moving beyond texture synthesis to complex scenes. The multi-scale pyramid of fully convolutional GANs effectively captures internal patch distributions at different scales, allowing diverse yet faithful sample generation. The authors acknowledge that semantic diversity is more limited compared to external methods trained on large datasets — the model cannot generate entirely new semantic concepts not present in the training image. Nevertheless, the broad utility for image manipulation tasks without additional training demonstrates the power of internal learning, suggesting that single-image generative models represent a promising and complementary paradigm to dataset-level generative approaches.