Infinite Texture introduces a novel approach for high-resolution texture generation using a text-to-image diffusion model. The process involves three main stages: generating a reference texture image from a text prompt, fine-tuning a diffusion model to learn the statistical distribution of the texture, and combining the outputs to synthesize a high-resolution texture. By leveraging diffusion models, the method can generate arbitrarily large, high-quality textures based on a reference texture image. The approach involves denoising small patches of the texture and aggregating them to synthesize the final high-resolution texture in a score aggregation manner.

To address challenges in texture synthesis, such as choosing the correct prompt and generating textures beyond 1024x1024 resolution, Infinite Texture utilizes a text-based interface for high-quality texture generation. By fine-tuning a diffusion model with a reference texture, the method can learn texture statistics and generate novel samples with stochastic permutations of the reference texture. The use of random crops for denoising patches improves runtime significantly while maintaining image quality. Additionally, the method introduces a strategy for generating spatially-consistent and large textures at inference time, overcoming memory constraints and enabling the synthesis of high-resolution textures seamlessly.