Score Distillation Sampling

Score Distillation Sampling (SDS) leverages pre-trained 2D image diffusion models to guide the optimization of 3D models, enabling text-to-3D generation and 3D editing. Current research focuses on mitigating common SDS limitations, such as over-smoothing, color saturation, and geometric inconsistencies, through refined loss functions, novel sampling strategies (e.g., using rectified flows or ODEs), and incorporating multi-view consistency. These advancements are significantly improving the fidelity and efficiency of text-to-3D generation, impacting fields like computer graphics, virtual reality, and potentially accelerating the creation of realistic 3D assets.