Research map / Image generation / Fast diffusion sampling
ProReflow: Progressive Reflow with Decomposed Velocity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Improves rectified-flow reflow for few-step text-to-image generation via progressive multi-stage window reflow and an aligned v-prediction loss that adds a cosine direction term to MSE, distilling SDv1.5/SDXL into 2–4 step models.
Problem
Reflow-based rectified flow for accelerating diffusion models into few-step (even one-step) generation is hard to optimize, since the teacher's velocity varies greatly across timesteps and magnitude-focused matching ignores trajectory direction.
Approach
Progressive reflow trains across windows in stages (8 to 4 to 2 windows) using cross-window reflow as curriculum, and aligned v-prediction adds a (1 - cosine similarity) direction term to the MSE velocity loss.
Key results
- ProReflow-II on SDv1.5 reaches FID 10.70 on COCO-2014 with 4 steps, close to its 32-step DDIM teacher (FID 10.05) and beating PeRFlow (12.01).
- On COCO-2017, ProReflow-II at 4 steps gets FID 22.03 versus 2-ReFlow's 32.97 at 4 steps.
- On SDXL with 4 steps, ProReflow-SDXL gets FID 25.36 on COCO2017 and 19.10 on COCO2014-10k, the best among compared methods in its tables.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add progressive reflow (ablation) | ProReflow-I · COCO-2017 validation | FID: 23.46 -> 22.97 | Helped |
| add progressive reflow (ablation) | ProReflow-I · COCO-2017 validation | CLIP score: 30.21 -> 30.29 | Helped |
| add aligned v-prediction (ablation) | ProReflow-I · COCO-2017 validation | FID: 23.09 -> 22.97 | Helped |
| add aligned v-prediction (ablation) | ProReflow-I · COCO-2017 validation | CLIP score: 30.25 -> 30.29 | Helped |
| add progressive reflow and aligned v-prediction (ablation) | ProReflow-I · COCO-2017 validation | FID: 23.81 -> 22.97 | Helped |
| add progressive reflow and aligned v-prediction (ablation) | ProReflow-I · COCO-2017 validation | CLIP score: 30.24 -> 30.29 | Helped |
| use ProReflow-II instead of PeRFlow at 4 sampling steps | Stable Diffusion v1.5 · COCO-2017 validation | FID: 23.81 -> 22.03 | Helped |
| use ProReflow-II instead of PeRFlow at 4 sampling steps | Stable Diffusion v1.5 · COCO-2014 validation | FID: 12.01 -> 10.70 | Helped |
| use ProReflow-SDXL instead of Rectified Diffusion | SDXL · COCO-2017 validation | FID: 25.81 -> 25.36 | Helped |
| use ProReflow-SDXL instead of DMDv2 | SDXL · COCO-2014-10k validation | FID: 19.32 -> 19.10 | Helped |
Surprising findings
- The paper reports that PeRFlow's FID increases when going from 4 to 8 sampling steps, and that progressive training removes this step-scalability problem.
- ProReflow-II, trained with only 2 windows, achieves lower FID at 4 steps than ProReflow-I trained with 4 windows.
Limitations
- One-step generation was not validated because the single-window model could not be trained to convergence under compute limits.
- The alpha hyperparameter was tuned with FID on COCO-30K, which is also a reported benchmark.
- The theoretical justification via VC-style bounds is heuristic and not verified empirically.
Experimental setup
Datasets: LAION-Art laion2B-en-aesthetic MS COCO 2014 MS COCO 2017
Benchmarks: MS COCO 2014 validation (FID-30K) MS COCO 2017 validation
Metrics: FID CLIP score
Baselines: 2-ReFlow InstaFlow-0.9B PeRFlow DPMSolver LCM-LoRA UniPC Flash Diffusion PCM Rectified Diffusion SDXL-Lightning SDXL-Turbo LCM
Builds on: Rectified Flow PeRFlow InstaFlow DDIM
Closest papers by meaning
- Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow (ICLR 2025 · similarity 0.92)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.91)
- Improving the Training of Rectified Flows (NeurIPS 2024 · similarity 0.91)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.89)
- Balanced Conic Rectified Flow (NeurIPS 2025 · similarity 0.88)
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · 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.