Research map / Image generation / Fast diffusion sampling
ParaFlow: Parallel Sampling for Flow Matching Models
Jianrong Lu, Bangwei Li, Haomin Zhang, Zhuoya Gu, Yongqing Lu, Jianhai Chen, Qinming He
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 velocity approximation (r=1) instead of full parallel evaluation (r=100) (ablation) | Euler + ParaFlow · SD3 100 steps | NFE / speedup: NFE 211.78 -> 28.34; speedup 3.3x -> 3.5x (single GPU vs 8 GPUs) | Improved |
| use velocity approximation (r=1) on Flux (ablation) | Euler + ParaFlow · Flux 100 steps | NFE / speedup: NFE 187.95 -> 22.93; speedup 4.0x -> 4.3x | Improved |
| apply ParaFlow (r=1) to sequential Euler on SD3 | Euler · SD3 100 steps | FID: 61.54 -> 61.96 | No clear effect |
| apply ParaFlow (r=1) to sequential Euler on SD3 | Euler · SD3 100 steps | CLIP score: 32.38 -> 32.17 | No clear effect |
| apply ParaFlow to Euler on Flux | Euler · Flux 100 steps | FID: 61.47 -> 62.28 | No clear effect |
| apply ParaFlow (r=1) to Euler across step counts | Euler · SD3 and Flux | wall-clock speedup: SD3 3.5x/2.8x/2.1x/1.6x and Flux 4.3x/3.6x/2.8x/1.7x for 100/75/50/25 steps | Improved |
| increase tolerance from 0.005 to 0.01 (ablation) | Euler + ParaFlow · SD3 25 steps | FID vs speedup: speedup 1.4x -> 1.7x; FID 64.11 -> 65.61 | Mixed |
| raise resolution from 512 to 2048 for ParaFlow r=1 (ablation) | Euler + ParaFlow · Flux 25 steps | speedup: 3.1x -> 1.4x | Hurt |
| reduce window size p from 32 to 8 (ablation) | Euler + ParaFlow · Flux 512 50 steps | speedup: 1.6x -> 2.8x; NFE 472.44 -> 123.51 | Improved |
| reduce window size p from 32 to 8 (ablation) | Euler + ParaFlow · Flux 512 50 steps | FID: 59.83 -> 59.82 | No clear effect |
| reduce parallel model evaluations r from 50 to 1 (ablation) | Euler + ParaFlow · Flux 512 50 steps | speedup: 1.3x -> 3.2x; NFE 472.44 -> 16.00 | Improved |
| reduce r from 50 to 1 (ablation) | Euler + ParaFlow · Flux 512 50 steps | FID: 59.94 -> 59.96 | No clear effect |
Closest papers by meaning
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.90)
- Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.89)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.89)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.89)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.88)
- Self-Refining Diffusion Samplers: Enabling Parallelization via Parareal Iterations (NeurIPS 2024 · similarity 0.88)
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.
