Research map / Image generation / Autoregressive and latent generation
Towards Bridging the Performance Gaps of Joint Energy-Based Models
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 |
|---|---|---|---|
| Add SAM to JEM (ablation) | JEM · CIFAR10 | accuracy: 89.5 -> 90.1 | Helped |
| Add SAM to JEM (ablation) | JEM · CIFAR10 | FID: 36.2 -> 35.0 | Helped |
| Add SAM to JEM++ (ablation) | JEM++ · CIFAR10 | accuracy: 93.5 -> 94.1 | Helped |
| Add SAM to JEM++ (ablation) | JEM++ · CIFAR10 | FID: 37.1 -> 36.6 | Helped |
| Remove DA from the likelihood pipeline (ablation) | JEM++ · CIFAR10 | FID: 37.1 -> 12.9 | Helped |
| Add SAM after removing DA from the likelihood pipeline (ablation) | JEM++ w/o DA · CIFAR10 | accuracy: 93.6 -> 96.0 | Helped |
| Add SAM after removing DA from the likelihood pipeline (ablation) | JEM++ w/o DA · CIFAR10 | FID: 12.9 -> 11.4 | Helped |
| Replace SAM with energy L2 regularization (ablation) | JEM++ w/o DA · CIFAR10 | accuracy: 96.0 -> 93.4 | Hurt |
| Use a larger SGLD sampling-step count K (ablation) | SADA-JEM · CIFAR10 | FID: 9.41 -> 11.4 -> 13.1 | Hurt |
| Use a larger SGLD sampling-step count K (ablation) | SADA-JEM · CIFAR10 | accuracy: 95.5 -> 96.0 -> 96.1 | Helped |
| Use SAM with rho=0.8 instead of rho=0.2 (ablation) | SADA-JEM · CIFAR10 | accuracy: 96.0 -> 91.9 | Hurt |
| Use SAM with rho=0.8 instead of rho=0.2 (ablation) | SADA-JEM · CIFAR10 | FID: 11.4 -> 19.5 | Hurt |
Closest papers by meaning
- Learning normalized image densities via dual score matching (NeurIPS 2025 · similarity 0.85)
- Generative Marginalization Models (ICML 2024 · similarity 0.85)
- Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling (NeurIPS 2025 · similarity 0.84)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.84)
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.84)
- GMAIL: Generative Modality Alignment for generated Image Learning (ICML 2025 · similarity 0.84)
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.