Research map / Probabilistic and scientific modeling / Variational and generative sampling
A Diffusive Classification Loss for Learning Energy-based Generative Models
RuiKang OuYang, Louis Grenioux, José Miguel Hernández-Lobato
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add the DiffCLF classification loss to DSM | VP diffusion model on MOG-40 · MOG-40, dim 8 to 128 | classification loss: 383.53 -> 4.40 (dim 128, DSM -> DiffCLF) | Improved |
| add the DiffCLF classification loss to DSM | VP diffusion model on MOG-40 · MOG-40 | Fisher divergence and MMD | No clear effect |
| use DiffCLF instead of conditional time score matching (CtSM) as the auxiliary loss to DSM (ablation) | VP diffusion model on MOG-40 · MOG-40, dim 8 | classification loss: 6.80 -> 4.41 | Improved |
| add DiffCLF to the stochastic interpolant energy training | SI between MOG-40 and MOG-2 · d=2 and d=128 log-density fit | log-density agreement with ground truth | Improved |
| train with DiffCLF in addition to DSM for molecular systems | energy-based diffusion models · Muller-Brown, Alanine Dipeptide, Chignolin | Langevin sample quality with learned energy | Improved |
| use DiffCLF instead of Fokker-Planck regularization (ablation) | energy-based diffusion models on molecules · molecular systems | training speed: 1.5-2.5x faster training at comparable performance | Improved |
| use DiffCLF-trained energies instead of DSM for SMC composition | OR and AND toy composition · toy compositions | composition accuracy | Improved |
| use DiffCLF-trained energies instead of DSM or CtSM in SMC Boltzmann generator | SI MOG-40 to MOG-2, 512-step SMC · Sliced Wasserstein and KS distance | W2 and KS distance | Improved |
| add Lclf to the Lbase objective for neural thermodynamic integration | Lbase (Mate et al., 2025) · ALDP solvation free energy | free energy estimate (reference 29.43): 27.30 -> 29.02 | Improved |
| increase number of classification levels N (ablation) | DiffCLF · theory (Corollary D.9) | asymptotic covariance norm | Improved |
Closest papers by meaning
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.88)
- Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling (NeurIPS 2025 · similarity 0.88)
- Improving Adversarial Energy-Based Model via Diffusion Process (ICML 2024 · similarity 0.87)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.87)
- Energy-Based Modelling for Discrete and Mixed Data via Heat Equations on Structured Spaces (NeurIPS 2024 · similarity 0.87)
- DisCo-Diff: Enhancing Continuous Diffusion Models with Discrete Latents (ICML 2024 · similarity 0.87)
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.
