Research map / Image generation / Fast diffusion sampling
Elucidating the Design Space of Diffusion-Based Generative 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 original ADM sampler with the proposed stochastic sampler and tune its settings | Pretrained ADM · ImageNet-64, class-conditional | FID: 2.07 to 1.55 | Helped |
| Retrain the ImageNet-64 model with the proposed training improvements | ADM architecture · ImageNet-64, class-conditional | FID: 1.55 to 1.36 | Helped |
| Add noise churn over the full noise range while keeping S_noise = 1 (ablation) | Pretrained diffusion models · CIFAR-10 VP and VE; ImageNet-64 | FID: 2.93 -> 2.69; 3.73 -> 2.97; 2.64 -> 1.86 | Helped |
| Allow stochasticity at all noise levels (ablation) | Pretrained diffusion models · CIFAR-10 VP and VE; ImageNet-64 | FID: 2.93 -> 2.54; 3.73 -> 2.51; 2.64 -> 1.63 | Helped |
| Set S_noise to 1 (ablation) | Pretrained diffusion models using the stochastic sampler · CIFAR-10 VP and VE; ImageNet-64 | FID: 2.54 -> 2.52; 2.51 -> 2.84; 1.63 -> 1.84 | Mixed |
| Use the proposed noise-level distribution and loss weighting with preconditioning (ablation) | Diffusion model training · CIFAR-10 and other evaluated training settings | FID | Helped |
| Add non-leaking data augmentation (ablation) | Diffusion model training · CIFAR-10 32×32 | FID: 1.85 to 1.79; 2.10 to 1.97 | Helped |
| Use deterministic rather than stochastic sampling (ablation) | Model trained with the proposed CIFAR-10 training setup · CIFAR-10 | FID | Helped |
| Use stochastic rather than deterministic sampling (ablation) | Newly trained class-conditional model · ImageNet-64 | FID | Helped |
| Use the proposed input, output, noise, and skip preconditioning (ablation) | Diffusion model training · Evaluated training configurations | FID | No clear effect |
| Use Heun's second-order method and the proposed time discretization (ablation) | Pretrained diffusion samplers · CIFAR-10 VP and VE; ImageNet-64 | FID: 2.85 -> 2.88; 5.45 -> 4.23; 2.85 -> 2.64 | Helped |
| Use the proposed noise schedule σ(t)=t and signal scaling s(t)=1 (ablation) | Pretrained VP and VE samplers · CIFAR-10 VP and VE | FID: 2.88 -> 2.93; 4.23 -> 3.73 | Mixed |
Closest papers by meaning
- Evaluating the design space of diffusion-based generative models (NeurIPS 2024 · similarity 0.91)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.90)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.90)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.90)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.90)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.90)
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.