Research map / Image generation / Fast diffusion sampling
Denoising Diffusion Implicit 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 |
|---|---|---|---|
| Use deterministic DDIM sampling (η=0.0) instead of DDPM sampling (η=1.0) with 10 steps | DDPM-trained image generator · CIFAR10 and CelebA | FID: 41.07 -> 13.36; 33.12 -> 17.33 | Helped |
| Use deterministic DDIM sampling (η=0.0) instead of DDPM sampling (η=1.0) with 100 steps | DDPM-trained image generator · CIFAR10 and CelebA | FID: 5.78 -> 4.16; 13.93 -> 6.53 | Helped |
| Use deterministic DDIM sampling (η=0.0) instead of DDPM sampling (η=1.0) with 1000 steps | DDPM-trained image generator · CIFAR10 and CelebA | FID: 4.73 -> 4.04; 5.98 -> 3.51 | Helped |
| Use deterministic DDIM sampling (η=0.0) instead of DDPM sampling (η=1.0) with 10 steps | DDPM-trained image generator · LSUN Bedroom and LSUN Church | FID: 42.78 -> 16.95; 51.56 -> 19.45 | Helped |
| Use deterministic DDIM sampling (η=0.0) instead of DDPM sampling (η=1.0) with 100 steps | DDPM-trained image generator · LSUN Bedroom and LSUN Church | FID: 6.81 -> 6.62; 8.27 -> 10.58 | Mixed |
| Increase the number of DDIM sampling steps (ablation) | DDIM · CIFAR10, CelebA, LSUN Bedroom, and LSUN Church | FID: 13.36 -> 4.04; 17.33 -> 3.51; 16.95 -> 6.62; 19.45 -> 10.58 | Helped |
| Increase stochasticity from η=0.0 to η=0.2 (ablation) | DDIM sampling · CIFAR10 (S=10) | FID: 13.36 -> 14.04 | Hurt |
| Increase stochasticity from η=0.0 to η=0.5 (ablation) | DDIM sampling · CIFAR10 (S=10) | FID: 13.36 -> 16.66 | Hurt |
| Increase stochasticity from η=0.0 to η=1.0 (ablation) | DDIM sampling · CIFAR10 and CelebA (S=10) | FID: 13.36 -> 41.07; 17.33 -> 33.12 | Hurt |
| Use the larger-variance σ sampling variant (ablation) | DDPM sampling · CIFAR10 and CelebA (S=10) | FID: 367.43; 299.71 | Hurt |
| Encode and decode with more DDIM steps (ablation) | DDIM · CIFAR10 test set | per-dimension mean squared error: 0.014 -> 0.0001 | Helped |
| Use quadratic rather than linear timestep selection (ablation) | DDIM sampling · CIFAR10 | FID | Helped |
Closest papers by meaning
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.94)
- UDPM: Upsampling Diffusion Probabilistic Models (NeurIPS 2024 · similarity 0.94)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.92)
- Learning to Discretize Denoising Diffusion ODEs (ICLR 2025 · similarity 0.92)
- Directly Denoising Diffusion Models (ICML 2024 · similarity 0.92)
- Denoising Diffusion Step-aware Models (ICLR 2024 · similarity 0.91)
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.