Research map / Probabilistic and scientific modeling / Variational and generative sampling
Adaptive Multi-stage Density Ratio Estimation for Learning Latent Space Energy-based Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace a simple latent Gaussian prior with the learned adaptive multi-stage latent EBM | generator model · SVHN, CelebA, CIFAR-10 | FID: 29.44 -> 26.19; 37.87 -> 35.38; 70.15 -> 65.01 | Helped |
| replace LEBM with adaptive multi-stage latent EBM training | generator model · SVHN, CelebA, CIFAR-10 | MSE: 0.008 -> 0.004; 0.013 -> 0.009; 0.020 -> 0.008 | Helped |
| replace LEBM with adaptive multi-stage latent EBM training | generator model · MNIST heldout digits 1, 4, 5, 7, 9 | AUPRC: 0.336± 0.01 -> 0.531 ± 0.02; 0.630± 0.02 -> 0.729 ± 0.02; 0.619± 0.01 -> 0.742 ± 0.01; 0.463 ± 0.01 -> 0.620 ± 0.02; 0.4 | Helped |
| add a learned latent EBM instead of using only the base Gaussian prior (ablation) | Adaptive CE · CelebA | FID: 62.78 -> 44.17 | Helped |
| add multiple density-ratio stages rather than use single-stage estimation (ablation) | Adaptive CE · CelebA | FID: 44.17 -> 35.38 | Helped |
| increase adaptive training from 4 to 8 density-ratio stages (ablation) | Adaptive CE · CelebA | FID: 35.38 -> 35.84 | Hurt |
| replace adaptive sequential training with non-adaptive training at 1 stage (ablation) | multi-stage density-ratio estimation · CelebA | FID: 44.17 -> 43.84 | Hurt |
| replace adaptive sequential training with non-adaptive training at 2 stages (ablation) | multi-stage density-ratio estimation · CelebA | FID: 39.85 -> 42.61 | Hurt |
| replace adaptive sequential training with non-adaptive training at 4 stages (ablation) | multi-stage density-ratio estimation · CelebA | FID: 35.38 -> 42.48 | Hurt |
| replace adaptive sequential training with non-adaptive training at 8 stages (ablation) | multi-stage density-ratio estimation · CelebA | FID: 35.84 -> 43.06 | Hurt |
| increase model size of a single-stage latent EBM to 4× (ablation) | single-stage latent EBM · CelebA | FID | No clear effect |
Closest papers by meaning
- Learning Energy-Based Models from Stochastic Interpolants using Spatiotemporal Differences (NeurIPS 2026 · similarity 0.87)
- Learning Latent Space Hierarchical EBM Diffusion Models (ICML 2024 · similarity 0.86)
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.85)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.85)
- Improving Adversarial Energy-Based Model via Diffusion Process (ICML 2024 · similarity 0.85)
- Density Ratio Estimation with Conditional Probability Paths (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.