DreamCraft3D is a novel method for generating high-quality 3D objects. The approach involves using a 2D reference image to guide the sculpting of geometry and enhancing textures. A key focus of this work is addressing the consistency issue that existing methods encounter. To address this, the authors use score distillation sampling and a view-dependent diffusion model. They also propose Bootstrapped Score Distillation (BSD) to specifically improve texture fidelity. The authors train a personalized diffusion model called Dreambooth by augmenting the renderings of the scene, enabling it to provide consistent guidance for 3D optimization. Through alternating optimization, they achieve mutually reinforcing improvements and generate coherent 3D objects with photorealistic renderings.
https://mrtornado24.github.io/DreamCraft3D/