Research map / Image generation / Fast diffusion sampling
Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic 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 |
|---|---|---|---|
| Replace the handcrafted reverse variance β_n with the analytic estimate | DDPM on CIFAR10 (LS) using OT · CIFAR10 (LS), K=1000 | NLL (bits/dim): 3.75 -> 3.59 | Helped |
| Replace the handcrafted reverse variance β_n with the analytic estimate | DDPM on CIFAR10 (CS) using OT · CIFAR10 (CS), K=1000 | NLL (bits/dim): 3.54 -> 3.42 | Helped |
| Replace the handcrafted reverse variance β_n with the analytic estimate | DDPM on CelebA 64x64 using OT · CelebA 64x64, K=1000 | NLL (bits/dim): 2.93 -> 2.66 | Helped |
| Replace the handcrafted reverse variance β_n with the analytic estimate | DDPM on ImageNet 64x64 using OT · ImageNet 64x64, K=4000 | NLL (bits/dim): 3.65 -> 3.61 | Helped |
| Replace the DDIM reverse variance with the analytic estimate | DDIM on CIFAR10 (LS) · CIFAR10 (LS), ET, K=50 | FID: 7.74 -> 4.04 | Helped |
| Replace the DDIM reverse variance with the analytic estimate | DDIM on CIFAR10 (CS) · CIFAR10 (CS), ET, K=50 | FID: 10.48 -> 6.02 | Helped |
| Replace the DDIM reverse variance with the analytic estimate | DDIM on CelebA 64x64 · CelebA 64x64, ET, K=50 | FID: 9.33 -> 6.13 | Helped |
| Replace the DDIM reverse variance with the analytic estimate | DDIM on ImageNet 64x64 · ImageNet 64x64, ET, K=100 | FID: 18.09 -> 17.73 | Helped |
| Use analytic reverse variance for DDPM sampling | DDPM on CIFAR10 (LS) · CIFAR10 (LS), ET, K=1000 | FID: 3.04 -> 4.03 | Hurt |
| Use analytic reverse variance for DDPM sampling | DDPM on CIFAR10 (CS) · CIFAR10 (CS), ET, K=400 | FID: 3.40 -> 3.96 | Hurt |
| Use analytic reverse variance for DDPM sampling | DDPM on CelebA 64x64 · CelebA 64x64, ET, K=1000 | FID: 3.16 -> 5.21 | Hurt |
| Use analytic reverse variance for DDPM sampling | DDPM on ImageNet 64x64 · ImageNet 64x64, ET, K=25 | FID: 29.21 -> 32.56 | Hurt |
Closest papers by meaning
- O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions (ICLR 2025 · similarity 0.91)
- Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.89)
- Reverse Transition Kernel: A Flexible Framework to Accelerate Diffusion Inference (NeurIPS 2024 · similarity 0.89)
- On Inference Stability for Diffusion Models (AAAI 2024 · similarity 0.89)
- Beyond Scores: Proximal Diffusion Models (NeurIPS 2025 · similarity 0.89)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.89)
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.