Research map / Image generation / Fast diffusion sampling
Straighten Viscous Rectified Flow via Noise Optimization
Jimin Dai, Jiexi Yan, Jian Yang, Lei Luo
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 |
|---|---|---|---|
| add historical velocity term (HVT) without noise optimization, 1 step (ablation) | Rectified flow (1-RF) · CIFAR-10 (NFE 1) | FID: 379 -> 332 | Improved |
| add HVT without noise optimization, 5 steps (ablation) | Rectified flow (1-RF) · CIFAR-10 (NFE 5) | FID: 34.81 -> 32.50 | Improved |
| add HVT without noise optimization, 10 steps (ablation) | Rectified flow (1-RF) · CIFAR-10 (NFE 10) | FID: 12.70 -> 9.34 | Improved |
| add noise optimization without HVT, 1 step (ablation) | Rectified flow (1-RF) · CIFAR-10 (NFE 1) | FID: 379 -> 4.72 | Improved |
| add noise optimization without HVT, 5 steps (ablation) | Rectified flow (1-RF) · CIFAR-10 (NFE 5) | FID: 34.81 -> 4.28 | Improved |
| add noise optimization without HVT, 10 steps (ablation) | Rectified flow (1-RF) · CIFAR-10 (NFE 10) | FID: 12.70 -> 4.75 | Improved |
| add HVT on top of noise optimization, 1 step (ablation) | VRFNO · CIFAR-10 (NFE 1) | FID: 4.72 -> 4.53 | Improved |
| add HVT on top of noise optimization, 5 steps (ablation) | VRFNO · CIFAR-10 (NFE 5) | FID: 4.28 -> 4.03 | Improved |
| add HVT on top of noise optimization, 10 steps (ablation) | VRFNO · CIFAR-10 (NFE 10) | FID: 4.75 -> 3.40 | Improved |
| replace 2-RF with VRFNO, AFHQ-CAT 64x64, 1 step | 2-RF · AFHQ-CAT 64x64 | FID: 181.93 -> 28.69 | Improved |
| replace 2-RF with VRFNO, AFHQ-DOG 256x256, 1 step | 2-RF · AFHQ-DOG 256x256 | FID: 189.82 -> 27.37 | Improved |
| replace CAF with VRFNO for flow straightness | CAF · CIFAR-10 | NFSS: 0.035 -> 0.026 | Improved |
Closest papers by meaning
- Balanced Conic Rectified Flow (NeurIPS 2025 · similarity 0.90)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.89)
- Improving the Training of Rectified Flows (NeurIPS 2024 · similarity 0.89)
- ProReflow: Progressive Reflow with Decomposed Velocity (CVPR 2025 · similarity 0.87)
- Coreset-Induced Conditional Velocity Flow Matching (NeurIPS 2026 · similarity 0.87)
- Variational Rectified Flow Matching (ICML 2025 · 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.
