핵심 개념
BNS solvers improve sample efficiency of diffusion and flow models by optimizing numerical solvers, achieving high PSNR and FID with low NFE.
초록
Introduces Bespoke Non-Stationary (BNS) Solvers for efficient sampling.
BNS solvers improve sample efficiency over existing methods.
Experimented with BNS solvers for image, text-to-image, and text-to-audio generation.
Detailed taxonomy of numerical solvers used for diffusion and flow models.
Demonstrated significant improvement in sample approximation with BNS solvers.
Compared BNS solvers with baselines and model distillation methods.
Conducted experiments on image, text-to-image, and text-to-audio generation tasks.
통계
BNS solver achieves 45 PSNR / 1.76 FID using 16 NFE in class-conditional ImageNet-64.
인용구
BNS solvers improve PSNR and FID over existing solvers.
BNS solvers demonstrate considerable improvement in sample approximation.