Provable Efficiency of Guidance in Diffusion Models for General Data Distribution

Authors: Gen Li, Yuchen Jiao

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present experimental results on the Gaussian Mixture Model (GMM) and Image Net dataset to demonstrate that guidance does not uniformly enhance the quality of all samples. Instead, it improves overall sample quality by reducing the average reciprocal of the classifier probability. This observation empirically validate our theoretical findings.
Researcher Affiliation Academia 1Department of Statistics, The Chinese University of Hong Kong, Hong Kong; Email: EMAIL. Correspondence to: Gen Li <EMAIL>.
Pseudocode No The paper describes mathematical equations for diffusion processes (e.g., (1a), (1b), (2a), (2b), (3a), (3b), (4), (5)) but does not present these or any other methodology in a structured pseudocode or algorithm block.
Open Source Code No The paper does not provide concrete access to source code for the methodology described. It mentions using a pre-trained diffusion model (Rombach et al., 2021) and the Inception v3 classifier (Szegedy et al., 2016), but these are third-party tools, not the authors' own implementation code.
Open Datasets Yes We conduct a numerical experiment on the Image Net dataset. Specifically, we generate samples using a pre-trained diffusion model (Rombach et al., 2021) with varying values of the guidance level w, and evaluate the classifier probabilities using the Inception v3 classifier (Szegedy et al., 2016).
Dataset Splits No For the Gaussian Mixture Model: "Specifically, we fix N = 4000, vary w from 0.01 to 10, and perform 104 trials for each w." For the Image Net dataset: "averaged over 20000 random trials — 20 trials for each of the 1000 Image Net categories." These details describe simulation/evaluation trials rather than training/validation/test splits of a dataset.
Hardware Specification No The paper does not provide any specific hardware details such as GPU/CPU models, memory specifications, or types of computing infrastructure used for the experiments.
Software Dependencies No The paper mentions using a "pre-trained diffusion model (Rombach et al., 2021)" and "Inception v3 classifier (Szegedy et al., 2016)" but does not specify any version numbers for these or other software dependencies.
Experiment Setup Yes Gaussian Mixture Model: Let us consider a distribution with two classes c = 0, 1, each with equal prior probability pc(0) = pc(1) = 0.5, in a one-dimensional data space (d = 1)... Specifically, we fix N = 4000, vary w from 0.01 to 10, and perform 104 trials for each w. Image Net dataset: Specifically, we generate samples using a pre-trained diffusion model (Rombach et al., 2021) with varying values of the guidance level w... averaged over 20000 random trials — 20 trials for each of the 1000 Image Net categories.