Research map / Image generation / Fast diffusion sampling
Antithetic Noise in Diffusion Models
Jing Jia, Sifan Liu, Bowen Song, Wei Yuan, Liyue Shen, Guanyang Wang
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 |
|---|---|---|---|
| use antithetic noise pairs (z, -z) instead of independent noise | unconditional diffusion on CelebA-HQ · CelebA-HQ | standard Pearson correlation: 0.25 -> -0.34 | Improved |
| centralize images before computing correlation of antithetic pairs | antithetic sampling on CelebA-HQ · CelebA-HQ | Pearson correlation: -0.34 -> -0.78 | Improved |
| use antithetic pairs on a VAE for MNIST | antithetic sampling · MNIST (VAE) | standard Pearson correlation: 0.21 vs 0.42 random; centralized -0.41 | Hurt |
| antithetic Monte Carlo (k=2) instead of standard MC | MC estimator · CelebA brightness CI | relative efficiency: MC 1.77 -> 0.26 CI length (47.41) | Improved |
| increase antithetic group size from k=2 to k=8 | AMC · CelebA brightness CI | relative efficiency: 47.41 -> 130.69 | Improved |
| increase antithetic group size from k=2 to k=8 | AMC · CIFAR-10 brightness CI | relative efficiency: 32.66 -> 30.35 | No clear effect |
| use antithetic MC instead of QMC | QMC · Stable Diffusion brightness CI | relative efficiency: 8.07 -> 18.53 | Improved |
| use antithetic MC instead of QMC | QMC · CIFAR-10 brightness CI | relative efficiency: 32.05 -> 32.66 | No clear effect |
| antithetic MC instead of standard MC for DPS inpainting | MC · DPS inpainting L1 | 95% CI width: 1.03 -> 0.83 | Improved |
| antithetic MC instead of standard MC for DPS super-resolution | MC · DPS super-resolution L1 | 95% CI width: 1.20 -> 1.01 | Improved |
| antithetic MC instead of standard MC for DPS Gaussian deblur PSNR | MC · DPS Gaussian deblur PSNR | 95% CI width: 0.29 -> 0.23 | Improved |
| antithetic MC instead of standard MC for DDS medical reconstruction | MC · DDS medical reconstruction L1 | relative efficiency: 1.44 for L1 and 1.37 for PSNR | Improved |
Closest papers by meaning
- Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.88)
- The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning (NeurIPS 2026 · similarity 0.88)
- Generalization through variance: how noise shapes inductive biases in diffusion models (ICLR 2025 · similarity 0.88)
- Beyond Scores: Proximal Diffusion Models (NeurIPS 2025 · similarity 0.87)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.87)
- Adapting to Unknown Low-Dimensional Structures in Score-Based 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.
