Score Distillation Sampling

Score Distillation Sampling (SDS) turns a pretrained diffusion model’s denoising scores into gradients that optimize another representation, such as a 3D scene, mesh, or drawing, without training a new diffusion model. Common approaches combine score-based guidance with multi-view gradient aggregation, geometry or pose regularization, reward models, and distribution-matching objectives to improve control and consistency. Work evaluates results in settings such as text-to-3D generation and editing, using measures of text alignment, visual quality, and multi-view consistency, while related distillation methods target faster generation and image or video compression. A central challenge is that noisy or mismatched score gradients can create artifacts, unstable optimization, or inconsistent views, so methods must balance fidelity, guidance strength, and computational cost.