Research map / Image generation / Fast diffusion sampling
Bi-Anchor Interpolation Solver for Accelerating Generative Modeling
Hongxu chen, Hongxiang Li, Zhen Wang, Long Chen
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 bi-anchor interpolation instead of single-anchor (add terminal backbone anchor) (ablation) | BA-solver · ImageNet-256, 7 NFE | FID: 4.35 -> 1.96 | Improved |
| add SideNet intermediate velocities (K=2) versus none (K=0) (ablation) | BA-solver · ImageNet-256, 7 NFE | FID: 3.54 -> 1.96 | Improved |
| use one intermediate velocity (K=1) instead of two (K=2) (ablation) | BA-solver · ImageNet-256, 7 NFE | FID: 1.96 -> 1.89 | No clear effect |
| use Gauss-Lobatto quadrature instead of Simpson rule (ablation) | BA-solver · ImageNet-256, 7 NFE | FID: 1.97 -> 1.96 | No clear effect |
| replace Flow-DPM with BA-solver | REPA-SiT backbone · ImageNet-256, 7 NFE | FID: 4.03 -> 1.96 | Improved |
| replace Flow-DPM with BA-solver | REPA-SiT backbone · ImageNet-512, 7 NFE | FID: 6.59 -> 2.88 | Improved |
| replace Euler with BA-solver at 5 NFE | REPA-SiT backbone · ImageNet-256, 5 NFE | FID: 11.58 -> 2.84 | Improved |
| replace Flow-DPM with BA-solver at 15 NFE | REPA-SiT backbone · ImageNet-256, 15 NFE | FID: 2.32 -> 1.65 | Improved |
| use BA-solver instead of Flow-DPM at 5 NFE | REPA-SiT backbone · ImageNet-256, 5 NFE | sFID: 7.04 -> 8.09 | Hurt |
| use BA-solver instead of Flow-DPM at 15 NFE | REPA-SiT backbone · ImageNet-256, 15 NFE | IS: 341.85 -> 326.67 | Hurt |
| train a 6M-parameter SideNet instead of fine-tuning the full backbone | BA-solver vs few-step methods · ImageNet-256 | training iterations: 313,000 (MeanFlow) -> 250 | Improved |
Closest papers by meaning
- Learning to Solve Generative ODEs Beyond the Linear Span (NeurIPS 2026 · similarity 0.90)
- Bespoke Solvers for Generative Flow Models (ICLR 2024 · similarity 0.88)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.88)
- Switched Flow Matching: Eliminating Singularities via Switching ODEs (ICML 2024 · similarity 0.87)
- Optimal Flow Matching: Learning Straight Trajectories in Just One Step (NeurIPS 2024 · similarity 0.87)
- On the Guidance of 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.
