Research map / Image generation / Fast diffusion sampling
Generalization and Memorization in Rectified Flow
Mingxing Rao, Daniel Moyer
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 |
|---|---|---|---|
| average N Monte Carlo noise samples (T_mc, #mc=5) instead of single-noise T_naive | membership inference on RF · CIFAR-10 | AUC: 70.98 -> 75.12 | Improved |
| T_mc instead of T_naive | membership inference on RF · SVHN | TPR@1%FPR: 1.89 -> 1.54 | Hurt |
| add input complexity calibration (divide by JPEG size) to T_mc (ablation) | T_mc · CIFAR-10 | TPR@1%FPR: 3.05 -> 27.88 | Improved |
| add input complexity calibration to T_mc (ablation) | T_mc · TinyImageNet | TPR@1%FPR: 5.33 -> 50.03 | Improved |
| add input complexity calibration to T_mc (ablation) | T_mc · SVHN | AUC: 70.92 -> 79.43 | Improved |
| add complexity calibration to T_naive (ablation) | T_naive · TinyImageNet | AUC: 72.33 -> 79.61 | Improved |
| add complexity calibration to T_naive (ablation) | T_naive · CIFAR-10 | TPR@1%FPR: 2.70 -> 3.70 | No clear effect |
| use JPEG instead of PNG for complexity measure (ablation) | T_mc_cal · CIFAR-10 | TPR: 11.84 -> 23.35 | Improved |
| use T_mc instead of T_naive across timesteps | membership inference · MSCOCO t=0.5 | AUC: 76.14 -> 85.33 | Improved |
| larger Monte Carlo sample count N_mc from 1 (ablation) | T_mc · CIFAR-10 | AUC | Improved |
| use Symmetric Exponential timestep sampling (alpha=2, 4) instead of uniform | Rectified Flow training · CIFAR-10, SVHN | peak MIA AUC and FID | Improved |
| non-identity data covariance (condition number kappa 1.0 to 100.0) | Rectified Flow toy Gaussian · Toy Gaussian | peak t: 0.50 -> 0.19 | Hurt |
Closest papers by meaning
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.87)
- Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training (NeurIPS 2025 · similarity 0.86)
- Gradient Variance Reveals Failure Modes in Flow-Based Generative Models (NeurIPS 2025 · similarity 0.86)
- Understanding and Mitigating Memorization in Generative Models via Sharpness of Probability Landscapes (ICML 2025 · similarity 0.85)
- The Geometric Inductive Bias of Grokking: Bypassing Phase Transitions via Architectural Topology (NeurIPS 2026 · similarity 0.85)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.85)
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.
