Text-to-Texture: Text-Guided 3D Mesh Texture Generation

We propose a zero-shot 3D mesh texture generation method based on text prompts. The proposed method builds upon and enhances the findings of existing research, with the primary objective being the utilization of a pre-trained latent diffusion model (LDM) for the generation of high-quality textures. In each denoising iteration, we denoise the renderings of the target 3D model in latent space and gradually adjust the texture to match the target text semantics through a differentiable renderer combined with backpropagation optimization. Our approach demonstrates efficacy across a range of test scenarios involving 3D models and diverse prompts, achieving varying degrees of optimization and enhancement in terms of generation quality, view consistency, and inference efficiency when compared to existing analogous techniques.