<ul><li>The paper introduces Uni-Instruct, a theory-driven framework that unifies over 10 existing one-step diffusion distillation approaches.</li><li>Uni-Instruct is inspired by a proposed diffusion expansion theory of the $f$-divergence family to effectively train one-step diffusion models.</li><li>Uni-Instruct achieves record-breaking Frechet Inception Distance (FID) values on benchmarks like CIFAR10 and ImageNet-$64	imes 64$, outperforming previous methods.</li><li>The application of Uni-Instruct on text-to-3D generation also shows improved generation quality and diversity compared to existing methods such as SDS and VSD.</li></ul>

Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction

Discover more