Mean-Shift Distillation for Diffusion Mode Seeking

TL;DR

Mean-shift distillation is a new diffusion technique that improves mode seeking and convergence in generative models, enhancing the quality of text-to-image and text-to-3D outputs without retraining.

Contribution

It introduces a novel diffusion distillation method based on mean-shift mode seeking, providing a provably accurate proxy for the distribution's gradient and a practical sampling procedure.

Findings

Superior mode alignment compared to existing methods

Improved convergence in synthetic and real-world tasks

Higher-fidelity results in text-to-image and text-to-3D applications

Abstract

We present mean-shift distillation, a novel diffusion distillation technique that provides a provably good proxy for the gradient of the diffusion output distribution. This is derived directly from mean-shift mode seeking on the distribution, and we show that its extrema are aligned with the modes. We further derive an efficient product distribution sampling procedure to evaluate the gradient. Our method is formulated as a drop-in replacement for score distillation sampling (SDS), requiring neither model retraining nor extensive modification of the sampling procedure. We show that it exhibits superior mode alignment as well as improved convergence in both synthetic and practical setups, yielding higher-fidelity results when applied to both text-to-image and text-to-3D applications with Stable Diffusion.