MOMENTUM-IMBUED LANGEVIN DYNAMICS (MILD) FOR FASTER SAMPLING

Nishanth Shetty*, Manikanta Bandla*, Nishit Neema*, Siddarth Asokan, Chandra Sekhar Seelamantula
Indian Institute of Science(IISc), Bengaluru
IEEE ICASSP'24
*Indicates Equal Contribution
Paper Code
MY ALT TEXT

A comparison of FID versus the number of function evaluations (NFEs) for the proposed MILD vis-a-vis the baseline ALD samplers, for varying number of sampling steps T. MILD sampler with T = 1 reduces the NFEs required to achieve a given FID by 2× to 5×.

Abstract

Score-based generative models have emerged as state-ofthe-art generative models. In this paper, we introduce a novel sampling scheme that can be combined with pretrained score-based diffusion models to speed up sampling by a factor of two to five in terms of the number of function evaluations (NFEs) with a superior Frechet Inception ´ distance (FID), compared to Annealed Langevin dynamics in noise-conditional score network (NCSN) and improved noise-conditional score network (NCSN++). The proposed sampling algorithm is inspired by momentum-based accelerated gradient descent used in convex optimization techniques. We validate the sampling efficiency of the proposed algorithm in terms of FID on CIFAR-10 and CelebA datasets.

Proposed Sampler

MY ALT TEXT

Schematic of the proposed accelerated Momentum-Imbued Langevin Dynamics (MILD) sampler.

GIF to show transition from noise to image

Poster

BibTeX

@inproceedings{shetty2024momentum,
        title={Momentum-Imbued Langevin Dynamics (MILD) for Faster Sampling},
        author={Shetty, Nishanth and Bandla, Manikanta and Neema, Nishit and Asokan, Siddarth and Seelamantula, Chandra Sekhar},
        booktitle={ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
        pages={6635--6639},
        year={2024},
        organization={IEEE}
      }