On the Trajectory Regularity of ODE-based Diffusion Sampling

Authors: Defang Chen, Zhenyu Zhou, Can Wang, Chunhua Shen, Siwei Lyu

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that trajectory regularity-based accelerated sampling can significantly improve the performance of diffusion-based generative models in a few (≤ 10) function evaluations.
Researcher Affiliation Collaboration 1State Key Laboratory of Blockchain and Data Security, Zhejiang University, China 2Hangzhou High-Tech Zone (Binjiang) Blockchain and Data Security Research Institute, China 3Zhejiang University, China 4University at Buffalo, USA.
Pseudocode Yes Algorithm 1 Geometry-Inspired Time Scheduling (standard dynamic programming, similar to (Watson et al., 2021))
Open Source Code No The paper states: 'The reported results of all compared approaches are obtained from an open-source toolbox: https://github.com/zju-pi/diff-sampler.' This refers to a toolbox for *compared approaches*, not explicitly the authors' own implementation of GITS.
Open Datasets Yes Table 1. Sample quality comparison in terms of Fréchet Inception Distance (FID (Heusel et al., 2017), lower is better) on four datasets (resolutions ranging from 32 × 32 to 256 × 256). CIFAR-10 32 × 32 (Krizhevsky & Hinton, 2009); FFHQ 64 × 64 (Karras et al., 2019); Image Net 64 × 64 (Russakovsky et al., 2015); LSUN Bedroom 256 × 256 (Yu et al., 2015).
Dataset Splits No The paper mentions datasets used for training but does not provide specific percentages or counts for training, validation, or test splits. It focuses on the generation aspect and FID scores of synthesized samples.
Hardware Specification Yes Table 4. Used time (seconds) in different stages of GITS, evaluated on an NVIDIA A100 GPU.
Software Dependencies No The paper mentions frameworks and methods like 'EDM framework', 'Euler method', 'i PNDM', 'DDIM', 'DPM-Solver', but does not provide specific version numbers for software libraries or dependencies used in their experiments.
Experiment Setup Yes The temporal domain is segmented using a polynomial function tn = (t0^(1/ρ) + n/N (tN^(1/ρ) - t0^(1/ρ)))^ρ, where t0 = 0.002, tN = 80, n ∈ [0, N], and ρ = 7. We initiate the dynamic programming experiments with 256 warmup samples...