VariedDreamer: Efficient Text-to-3D Generation via Optimizing Fine Geometric

The generation of 3D objects has significantly impacted healthcare, AR, and VR by enhancing understanding of complex structures. However, non-professionals often struggle with generating 3D objects, driving the development of Text-to-3D. Existing methods, including generation-based and optimizationbased approaches, have limitations such as poor quality for simple inputs and time-consuming optimization. This paper presents VariedDreamer, a novel framework with three stages: semanticenhanced 3D prior generation, mesh geometry polishing, and detailed texture amelioration. VariedDreamer leverages text-toimage models, multi-view diffusion models, mesh reconstruction, mesh polishing, and texture optimization techniques to generate high-quality 3D objects. It outperforms generation-based methods in understanding complex text and quality, and accelerates optimization compared to optimization-based methods while resolving the Janus problem.