Research map / Image generation / Fast diffusion sampling
Understanding Diffusion Objectives as the ELBO with Simple Data Augmentation
Diederik P Kingma, Ruiqi Gao
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 |
|---|---|---|---|
| Replace the Simple Diffusion weighting with shifted EDM-monotonic weighting and use the adaptive training schedule | Simple Diffusion U-ViT, L · ImageNet 128x128, no guidance | eval FID: 3.23 -> 2.88 | Improved |
| Use shifted sigmoid(-λ + 2) weighting and the adaptive training schedule instead of the baseline weighting and schedule | Simple Diffusion U-ViT, L · ImageNet 128x128, no guidance | Inception Score: 171.9 ± 2.46 -> 183.1 ± 2.20 | Improved |
| Use EDM-monotonic weighting and the adaptive training schedule | Simple Diffusion U-ViT, L · ImageNet 128x128, guidance | eval FID: 3.57 -> 3.16 | Improved |
| Use EDM-monotonic weighting and the adaptive training schedule | Simple Diffusion U-ViT, L · ImageNet 256x256, no guidance | eval FID: 3.75 -> 3.36 | Improved |
| Use EDM-monotonic weighting and the adaptive training schedule | Simple Diffusion U-ViT, L · ImageNet 512x512, no guidance | eval FID: 4.53 -> 4.09 | Improved |
| Replace the non-monotonic sech(λ/2) weighting with sigmoid(-λ + 2) | ε-prediction U-Net · ImageNet 64x64, DDPM sampler | FID: 1.85 -> 1.68 | Improved |
| Use sigmoid(-λ + 5) instead of sech(λ/2) | ε-prediction U-Net · ImageNet 64x64, DDPM sampler | FID: 1.85 -> 1.94 | Hurt |
| Replace the EDM training schedule with the adaptive schedule (ablation) | EDM parameterization · ImageNet 64x64, EDM sampler | FID: 1.45 -> 1.43 | Improved |
| Replace EDM weighting with sigmoid(-λ + 2) weighting (ablation) | EDM parameterization with adaptive schedule · ImageNet 64x64, EDM sampler | FID: 1.43 -> 1.55 | Hurt |
| Replace the original v-prediction weighting exp(-λ/2) with EDM-monotonic weighting (ablation) | v-prediction U-Net with adaptive schedule · ImageNet 64x64, EDM sampler | FID: 1.62 -> 1.45 | Improved |
| Replace the baseline exp(-λ/2) weighting with shifted EDM-monotonic weighting (ablation) | Simple Diffusion U-ViT, L · ImageNet 128x128, no guidance | Inception Score: 171.9 ± 2.46 -> 171.1 ± 2.67 | No clear effect |
Closest papers by meaning
- Diffusion Models With Learned Adaptive Noise (NeurIPS 2024 · similarity 0.90)
- DiffEnc: Variational Diffusion with a Learned Encoder (ICLR 2024 · similarity 0.88)
- Analyzing and Improving the Training Dynamics of Diffusion Models (CVPR 2024 · similarity 0.87)
- DiffPTS: Rethinking Diffusion ELBO for Probabilistic Time Series Forecasting (NeurIPS 2026 · similarity 0.87)
- Information Theoretic Learning for Diffusion Models with Warm Start (NeurIPS 2025 · similarity 0.87)
- Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models (NeurIPS 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.
