Research map / Image generation / Fast diffusion sampling
Learning Energy-Based Models by Diffusion Recovery Likelihood
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 |
|---|---|---|---|
| Use diffusion recovery likelihood (T6) instead of one time step (ablation) | Explicit EBM image generation · CIFAR-10 | FID: 32.12 -> 9.58 | Helped |
| Use diffusion recovery likelihood (T6) instead of one time step (ablation) | Explicit EBM image generation · CIFAR-10 | Inception Score: 6.72 ± 0.12 -> 8.30 ± 0.11 | Helped |
| Use 1000 time steps with the recovery-likelihood objective rather than the T6 setting (ablation) | Diffusion recovery likelihood · CIFAR-10 | FID: 9.58 -> 22.58 | Hurt |
| Use 1000 time steps with the recovery-likelihood objective rather than the T6 setting (ablation) | Diffusion recovery likelihood · CIFAR-10 | Inception Score: 8.30 ± 0.11 -> 7.71 ± 0.08 | Hurt |
| Train with the DSM objective instead of the recovery-likelihood objective at T = 1000 (ablation) | Diffusion recovery likelihood · CIFAR-10 | FID: 22.58 -> 21.76 | No clear effect |
| Train with the DSM objective instead of the recovery-likelihood objective at T = 1000 (ablation) | Diffusion recovery likelihood · CIFAR-10 | Inception Score: 7.71 ± 0.08 -> 7.76 ± 0.11 | No clear effect |
| Increase sampling steps from K = 30 to K = 50 (ablation) | Diffusion recovery likelihood (T = 6) · CIFAR-10 | FID: 9.58 -> 9.36 | Helped |
| Increase sampling steps from K = 30 to K = 50 (ablation) | Diffusion recovery likelihood (T = 6) · CIFAR-10 | Inception Score: 8.30 ± 0.11 -> 8.46 ± 0.13 | Helped |
| Use the proposed T6 method rather than COCO-GAN | Image generation · CelebA 64×64 | FID: 4.0 -> 5.98 | Hurt |
| Use the proposed T1k method rather than DistAug | Image density modeling · CIFAR-10 test | Bits per dimension: 2.56 -> 3.18 | Hurt |
| Use 100k long-run sampling steps rather than 1k sampling steps | T1k conditional MCMC · CIFAR-10 samples | FID: 25.12 -> 24.89 | No clear effect |
Closest papers by meaning
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.96)
- Improving Adversarial Energy-Based Model via Diffusion Process (ICML 2024 · similarity 0.90)
- Diffusion Models With Learned Adaptive Noise (NeurIPS 2024 · similarity 0.88)
- Outsourced Diffusion Sampling: Efficient Posterior Inference in Latent Spaces of Generative Models (ICML 2025 · similarity 0.88)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.88)
- Information Theoretic Learning for Diffusion Models with Warm Start (NeurIPS 2025 · similarity 0.88)
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.