Research map / Image generation / Fast diffusion sampling
PeRFlow: Piecewise Rectified Flow as Universal Plug-and-Play Accelerator
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.
Proposes PeRFlow, which splits the diffusion PF-ODE trajectory into time windows and applies reflow within each window using real data, with epsilon/velocity parameterizations initialized from pretrained Stable Diffusion models, yielding few-step generators and transferable weight deltas.
Problem
Few-step sampling of pretrained diffusion models is slow; existing reflow-based acceleration (InstaFlow) needs expensive synthetic ODE-generated datasets that limit quality and scalability.
Approach
Apply reflow piecewise over K time windows, sampling window starting points from real data via the forward noising process and solving only short-interval endpoints online, with parameterizations derived from epsilon-prediction so the student initializes from the teacher; the weight delta can be plugged into other SD workflows.
Key results
- On SD-v1.5, PeRFlow 4-step FID on LAION-5B is 8.60 vs. 15.28 for LCM-LoRA and 14.32 for InstaFlow.
- On SDXL, PeRFlow 4-step FID on COCO2014 is 18.48 vs. 19.74 for LCM-LoRA and 22.86 for Lightning.
- PeRFlow delta-W plugged into ControlNet-Tile enables 4-step 1024x1024 enhancement and Wonder3D enables one-step multiview generation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PeRFlow instead of InstaFlow | SD-v1.5 · LAION-5B, 4-step | FID: 14.32 -> 8.60 | Helped |
| Use PeRFlow instead of InstaFlow | SD-v1.5 · COCO2014, 4-step | FID: 13.86 -> 11.31 | Helped |
| Use PeRFlow instead of LCM-LoRA | SD-v1.5 · SD-v1.5 reference, 8-step | FID: 21.19 -> 5.03 | Helped |
| Use PeRFlow instead of Lightning | SDXL · LAION-5B, 4-step | FID: 15.47 -> 13.30 | Helped |
| Use PeRFlow instead of LCM-LoRA | SDXL · COCO2014, 4-step | FID: 19.74 -> 18.48 | Helped |
| Use PeRFlow instead of LCM-LoRA | SDXL · SDXL reference, 8-step | FID: 9.90 -> 9.12 | Helped |
| Increase inference steps from 4 to 8 | PeRFlow SD-v1.5 · COCO2014 | FID: 11.31 -> 14.16 | Hurt |
| Initialize the student from the DDIM update-rule parameterization [B] | PeRFlow · Training convergence | training convergence speed | Helped |
| Use CFG-fixed instead of CFG-sync | PeRFlow · Text-to-image generation | sampling diversity | Mixed |
Surprising findings
- LCM-LoRA FID increases from 4 to 8 steps on several benchmarks, whereas PeRFlow FID generally decreases.
Limitations
- Currently uses 4 time windows and needs at least 4 steps; 1-2 step inference not supported.
- Complex structures may require extra steps, with inference budget allocation needed for best quality.
- CFG-fixed trades sampling diversity for fewer failure cases.
Experimental setup
Datasets: LAION-Aesthetics-5+ MS COCO 2014 JourneyDB LAION-5B-Aesthetics
Benchmarks: LAION-5B-Aesthetics FID MS COCO 2014 FID SD-v1.5 generated FID
Metrics: FID
Baselines: InstaFlow LCM-LoRA SDXL-Lightning Progressive Distillation DDIM
Builds on: rectified flow InstaFlow DDIM DPM-Solver classifier-free guidance
Closest papers by meaning
- Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow (ICLR 2025 · similarity 0.89)
- Improving the Training of Rectified Flows (NeurIPS 2024 · similarity 0.88)
- ProReflow: Progressive Reflow with Decomposed Velocity (CVPR 2025 · similarity 0.87)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.87)
- RayFlow: Instance-Aware Diffusion Acceleration via Adaptive Flow Trajectories (CVPR 2025 · similarity 0.87)
- Text-to-Image Rectified Flow as Plug-and-Play Priors (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.