Research map / Image generation / Fast diffusion sampling
Rényi Diffusion Models
Yirong Shen, Lu GAN, Cong Ling
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 KL-based ELBO with Renyi objective (alpha=0.5) (ablation) | eps-prediction diffusion, cosine schedule · CIFAR-10 | NLL (bits/dim): 2.74 -> 2.48 | Improved |
| replace KL-based ELBO with Renyi objective (ablation) | eps-prediction diffusion, cosine schedule · ImageNet-64 | NLL (bits/dim): 3.57 -> 3.29 | Improved |
| replace KL with Renyi objective without weighted loss (ablation) | eps-prediction diffusion · CIFAR-10 | FID: 6.31 -> 6.33 (DDPM sampler) | No clear effect |
| reduce Renyi order alpha from 1 toward 0.5 (ablation) | Renyi diffusion · CIFAR-10 | NLL (bits/dim): 2.74 -> 2.48 | Improved |
| reduce alpha to very small values (0.0001) (ablation) | Renyi diffusion · CIFAR-10 | FID: 6.06 -> 28.79 (NLL 2.74 -> 2.29) | Hurt |
| add time-dependent weighted loss to KL baseline (ablation) | eps-prediction baseline · CIFAR-10 | FID / NLL: FID 6.31 -> 2.53; NLL 2.74 -> 3.97 | Mixed |
| add weighted loss to Renyi objective (ablation) | Renyi (eps) · CIFAR-10 | FID / NLL: FID 6.33 -> 2.51; NLL 2.48 -> 2.66 | Mixed |
| add importance sampling with Fourier features to Renyi + weighted loss (ablation) | Renyi (eps) + WL · CIFAR-10 | FID / NLL: FID 2.51 -> 2.27; NLL 2.66 -> 2.24 | Improved |
| replace cosine with adaptive noise schedule for Renyi (eps) (ablation) | Renyi (eps) · CIFAR-10 | NLL (bits/dim): 2.13 -> 2.10 | Improved |
| use v-prediction with Renyi objective (ablation) | Renyi diffusion · CIFAR-10 | NLL (bits/dim): 2.10 -> 2.04 | Improved |
| train with Renyi bound instead of separate FID- and NLL-favorable regimes | RDM · CIFAR-10 | NLL and FID: NLL 2.09, FID 2.08 | Improved |
| Renyi tighter bound shortens training | RDM · training iterations | iterations to converge: approx 770k vs 2M (i-DODE, W-PCDM) and 10M (VDM) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- PaGoDA — PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher (NeurIPS 2024) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Training Diffusion-based Generative Models with Limited Data (ICML 2025 · similarity 0.89)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.89)
- Evaluating the design space of diffusion-based generative models (NeurIPS 2024 · similarity 0.89)
- Nearest Neighbour Score Estimators for Diffusion Generative Models (ICML 2024 · similarity 0.88)
- Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion (ICML 2025 · similarity 0.88)
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · 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.
