Abstract
We present Infinite Texture, a method for generating arbitrarily largetexture images from a text prompt. Our approach fine-tunes a diffusion model ona single texture, and learns to embed that statistical distribution in theoutput domain of the model. We seed this fine-tuning process with a sampletexture patch, which can be optionally generated from a text-to-image modellike DALL-E 2. At generation time, our fine-tuned diffusion model is usedthrough a score aggregation strategy to generate output texture images ofarbitrary resolution on a single GPU. We compare synthesized textures from ourmethod to existing work in patch-based and deep learning texture synthesismethods. We also showcase two applications of our generated textures in 3Drendering and texture transfer.