Layout-to-Image (L2I) was proposed by a research team at the University of British Columbia in November 2018, and the relevant research results were published in the paper "IImage Generation from Layout", selected for CVPR 2019.

Layout-to-Image (L2I) is a novel layout-based image generation method. Given a coarse spatial layout (bounding boxes + object categories), the model can generate a set of realistic images with the correct objects in ideal positions. L2I decomposes the representation of each object into deterministic category information and uncertain appearance information. The category is encoded using word embeddings, while the appearance is mapped to a low-dimensional latent vector sampled from a normal distribution. The object representations are then ensembled by a convolutional LSTM, converging into a unified encoding of the entire layout, and finally decoded into a complete image.