Learning Latent Space Hierarchical EBM Diffusion Models
Authors: Jiali Cui, Tian Han
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive experiments demonstrate a superior performance of our diffusion-learned EBM prior on various challenging tasks. We assess our model on the standard benchmark CIFAR-10 and the challenging high-resolution Celeb A-HQ-256 and large-scale LSUN-Church-64. We recruit Fr e Chet Inception Distance (FID) and Inception Score (IS) metrics to evaluate the quality of image synthesis. We report our results in Tab. 1 and Tab. 2 as well as the FID score of the reconstructed images. More quantitative and qualitative results can be found in the ablation studies and App. B. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Stevens Institute of Technology. Correspondence to: Tian Han <than6@stevens.edu>. |
| Pseudocode | Yes | Algorithm 1 Learning EBM parameter ω; Algorithm 2 Sampling and Image Synthesis |
| Open Source Code | Yes | Our project page is available at https://jcui1224.github.io/diffusion-hierarchical-ebm-proj/. |
| Open Datasets | Yes | We assess our model on the standard benchmark CIFAR-10 and the challenging high-resolution Celeb A-HQ-256 and large-scale LSUN-Church-64. |
| Dataset Splits | No | The paper mentions CIFAR-10, Celeb A-HQ-256, and LSUN-Church-64 datasets but does not explicitly state the training, validation, and test splits (e.g., percentages or specific counts for each split). |
| Hardware Specification | No | No specific hardware (GPU models, CPU models, memory) used for the experiments is mentioned in the paper. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | We conduct such experiments to demonstrate a smooth energy landscape learned for our EBMs prior. Diffusion step T. First, we train our diffusion-based EBM prior with more diffusion steps, e.g., T = 6. Langevin step K. By using more Langevin steps, we should explore the energy landscape better and obtain more effective EBM samples for learning. We show our results in Tab. 3 where using 50 steps (denoted as K = 50) delivers better synthesis than using 30 steps, while using 100 steps only shows a minor improvement but costs much more training and sampling time. |