Abstract
Score Distillation Sampling (SDS) has made significant strides in distillingimage-generative models for 3D generation. However, itsmaximum-likelihood-seeking behavior often leads to degraded visual quality anddiversity, limiting its effectiveness in 3D applications. In this work, wepropose Consistent Flow Distillation (CFD), which addresses these limitations.We begin by leveraging the gradient of the diffusion ODE or SDE samplingprocess to guide the 3D generation. From the gradient-based samplingperspective, we find that the consistency of 2D image flows across differentviewpoints is important for high-quality 3D generation. To achieve this, weintroduce multi-view consistent Gaussian noise on the 3D object, which can berendered from various viewpoints to compute the flow gradient. Our experimentsdemonstrate that CFD, through consistent flows, significantly outperformsprevious methods in text-to-3D generation.