Research map / Probabilistic and scientific modeling / Variational and generative sampling
Improving Rectified Flow with Boundary Conditions
Xixi Hu, Runlong Liao, Keyang Xu, Bo Liu, Yeqing Li, Eugene Ie, Hongliang Fei, Qiang Liu
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 |
|---|---|---|---|
| enforce both t=0 and t=1 boundary conditions via mask-based parameterization (ablation) | Rectified Flow DiT-B/2 · ImageNet 256x256 (SDE) | FID: 6.79 -> 6.18 | Improved |
| enforce only the right boundary (t=1) with a single pass (ablation) | Rectified Flow DiT-B/2 · ImageNet 256x256 (Euler) | FID: 6.87 -> 6.80 | Improved |
| right-boundary-only enforcement lowers IS (ablation) | Rectified Flow DiT-B/2 · ImageNet 256x256 (Euler) | IS: 130.08 -> 125.44 | Hurt |
| use subtraction-based parameterization (extra forward pass at t=1) (ablation) | Rectified Flow DiT-B/2 · ImageNet 256x256 (Euler) | FID: 6.87 -> 6.32 | Improved |
| boundary enforcement versus mode sampling of timesteps near boundaries | Rectified Flow DiT-B/2 · ImageNet 256x256 (Euler) | FID: 7.48 -> 6.32 | Improved |
| add classifier-free guidance with SDE sampler | Subtraction-based Boundary RF · ImageNet 256x256 with CFG (SDE) | FID: 7.20 -> 6.94 vs vanilla; but CFG SDE worse than no-CFG 6.38 | Mixed |
| use cosine boundary functions instead of offset cosine, quadratic, sqrt, or linear (ablation) | Mask-based Boundary RF · ImageNet 256x256 | FID and IS | Improved |
| enforce boundary conditions under stronger stochasticity (overshooting sampler noise scale c=5) | Mask-based Boundary RF · ImageNet 256 overshooting sampler c=5 | FID: 11.76 -> 10.68 | Improved |
| scale boundary RF to larger models and DiT-L SDE sampler | Mask-based Boundary RF · ImageNet 256x256 (SDE) DiT-L | FID: 6.77 -> 4.86 | Improved |
| mask-based boundary on DiT-L with Euler sampler | Mask-based Boundary RF · ImageNet 256x256 (Euler) DiT-L | IS: 165.44 -> 162.34 | Hurt |
Closest papers by meaning
- Improving the Training of Rectified Flows (NeurIPS 2024 · similarity 0.90)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.88)
- Balanced Conic Rectified Flow (NeurIPS 2025 · similarity 0.88)
- Text-to-Image Rectified Flow as Plug-and-Play Priors (ICLR 2025 · similarity 0.87)
- Towards Hierarchical Rectified Flow (ICLR 2025 · similarity 0.87)
- Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow (ICLR 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.
