Provable Maximum Entropy Manifold Exploration via Diffusion Models

Authors: Riccardo De Santi, Marin Vlastelica, Ya-Ping Hsieh, Zebang Shen, Niao He, Andreas Krause

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we empirically evaluate our approach on both synthetic and high-dimensional text-to-image diffusion, demonstrating promising results.
Researcher Affiliation Academia 1ETH Zurich, 8092 Zurich, Switzerland 2ETH AI Center, Zurich, Switzerland. Correspondence to: Riccardo De Santi <EMAIL>.
Pseudocode Yes Algorithm 1 Score-based Maximum Entropy Manifold Exploration (S-MEME) Algorithm 2 LINEARFINETUNINGSOLVER (Implementation based on Adjoint Matching (Domingo-Enrich et al., 2024))
Open Source Code No The paper does not provide an explicit statement or link for open-source code availability for the methodology described.
Open Datasets Yes For this we utilize the stable diffusion (SD) 1.5 (Rombach et al., 2021) checkpoint pre-trained on the LAION-5B dataset (Schuhmann et al., 2022).
Dataset Splits No The paper uses pre-trained models and discusses fine-tuning and sampling for evaluation. It mentions pre-training on 10K samples for an illustrative setting, but does not specify dataset splits (e.g., train/test/validation) for reproducing their experiments.
Hardware Specification Yes We fine-tuned the checkpoint with K = 3 iterations of S-MEME on a single Nvidia H100 GPU for the prompt A creative architecture.
Software Dependencies No The paper mentions using "stable diffusion (SD) 1.5" but does not specify version numbers for other key software components like programming languages (e.g., Python) or libraries (e.g., PyTorch, TensorFlow).
Experiment Setup Yes For fine-tuning, in this experiment we ran S-MEME for 6000 gradient steps in total, for K = 1, 2, 3, 4. ... we perform an iteration of Algorithm 2 by first sampling 20 trajectories via DDPM of length 400 that are used for solving the lean adjoint ODE with the reward λ log p T (x) and λ = 0.1. Subsequently we perform 2 stochastic gradient steps by the Adam optimizer with batch size 2048, initialized with learning rate 4 10 4.