This paper points out that although recent texture generation methods have achieved remarkable results thanks to the powerful generative dictionary exploited in large-scale text-to-image diffusion models, abstract text prompts have limitations in providing global texture or shape information, resulting in blurry or inconsistent patterns. To address this issue, this paper presents FlexiTex, which embeds rich information through visual guidance to generate high-quality textures. The core of FlexiTex is the Visual Guidance Enhancement module, which reduces the ambiguity of text prompts by incorporating more specific information in the visual guidance and preserves high-frequency details. To further enhance the visual guidance, this paper introduces the Direction-Aware Adaptation module, which automatically designs direction prompts according to different camera poses to avoid the Janus problem and maintain semantic global consistency. With the advantage of visual guidance, FlexiTex produces quantitatively and qualitatively excellent results, showing the potential to advance texture generation for real-world applications.
