Maximum Likelihood Training of Implicit Nonlinear Diffusion Model

Authors: Dongjun Kim, Byeonghu Na, Se Jung Kwon, Dongsoo Lee, Wanmo Kang, Il-chul Moon

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, INDM achieves the state-of-the-art FID of 1.75 on Celeb A. We release our code at https://github.com/byeonghu-na/INDM.
Researcher Affiliation Collaboration Dongjun Kim KAIST dongjoun57@kaist.ac.kr Byeonghu Na KAIST wp03052@kaist.ac.kr Se Jung Kwon NAVER CLOVA Dongsoo Lee NAVER CLOVA Wanmo Kang KAIST Il-Chul Moon KAIST / Summary.AI
Pseudocode Yes Algorithm 1 Implicit Nonlinear Diffusion Model
Open Source Code Yes We release our code at https://github.com/byeonghu-na/INDM.
Open Datasets Yes This section quantitatively analyzes suggested INDM on CIFAR-10 and Celeb A 64 64.
Dataset Splits No We compute NLL/NELBO for performances of density estimation with Bits Per Dimension (BPD).
Hardware Specification No The paper acknowledges support from government grants but does not specify any particular hardware components such as GPU or CPU models used for the experiments.
Software Dependencies No Throughout the experiments, we use NCSN++ with VESDE and DDPM++ with VPSDE [1] as the backbones of diffusion models, and a Res Net-based flow model [23, 24] as the backbone of the flow model.
Experiment Setup Yes See Appendix F for experimental details. We experiment with a pair of weighting functions for the score training. One is the likelihood weighting [11] with λ(t) = g2(t)... The other is the variance weighting [8] λ(t) = σ2(t)... We use either the Predictor-Corrector (PC) sampler [1] or a numerical ODE solver (RK45 [25])... For a better FID, we find the optimal signal-to-noise value (Table 14), sampling temperature (Table 15), and stopping time (Table 16).