SinGAN

SinGAN is a generative model designed to learn a generative distribution from a single natural image. Unlike conventional GANs that require large datasets, SinGAN builds a multi-scale, fully convolutional architecture that captures the internal statistics of the single input image. The model consists of a cascade of generators and discriminators trained at progressively higher resolutions.

At each scale, the network learns to generate image patches that resemble the corresponding patches in the input image, conditioned on noise inputs. Training proceeds in a bottom-up fashion, starting from the smallest scale and moving to larger scales, ensuring that coarse structure is established before adding fine details. After training, new image samples can be produced by sampling noise vectors for each scale, enabling diverse textures and structures that resemble the input.

Applications include texture synthesis, where new samples share the input's texture; image editing and manipulation, such as local inpainting, content-aware synthesis, and style transfer within a single image family; and super-resolution or upscaling with texture-aware detail preservation. The key advantage is that no external data is required; the method leverages the redundancy of patches within a natural image.

Limitations include reliance on the input image's internal statistics, which means the method works best for images with strong repetitive patterns or textures. It may struggle with images dominated by global structure or unique content. Training can be slow for high-resolution images and requires careful choice of scales, patch sizes, and regularization.

SinGAN has inspired further single-image generative approaches and variants that extend its ideas to conditional or multi-image training.