Research map / Image generation / Fast diffusion sampling
pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation
Hansheng Chen, Kai Zhang, Hao Tan, Leonidas Guibas, Gordon Wetzstein, Sai Bi
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 GM policy instead of DX policy (ablation) | pi-Flow (REPA teacher) · ImageNet 256 (1-NFE), 40K iterations | FID: 4.44 -> 3.07 | Improved |
| add GM dropout 0.05 (ablation) | pi-Flow GM K=32 (FM teacher) · ImageNet 256 (1-NFE) | FID: 4.14 -> 3.65 | Improved |
| increase GM components K from 8 to 32 (ablation) | pi-Flow GM (REPA teacher) · ImageNet 256 (1-NFE), 40K iterations | FID: 3.07 -> 3.08 | No clear effect |
| vary DX grid points N from 20 to 40 (ablation) | pi-Flow DX policy · ImageNet 256 (1-NFE), 40K iterations | FID: 4.44 -> 4.90 | Hurt |
| use REPA teacher instead of FM teacher | pi-Flow GM K=32 · ImageNet 256 (1-NFE) | FID: 3.65 -> 3.08 | Improved |
| add GM temperature (ablation) | pi-Flow (GM-FLUX) 4-NFE · HPSv2 prompts, teacher-referenced | FID: 14.9 -> 14.3 | Improved |
| add micro-window velocity matching (ablation) | pi-Flow (GM-FLUX) 4-NFE · HPSv2 prompts, teacher-referenced | FID: 14.6 -> 14.3 | Improved |
| use pi-ID instead of VSD distillation for diversity | Qwen-Image Lightning · OneIG-Bench (Qwen-Image, 4-NFE) | Diversity: 0.116 -> 0.180 | Improved |
| use pi-ID instead of VSD distillation for diversity | SenseFlow (FLUX) · OneIG-Bench (FLUX, 4-NFE) | Diversity: 0.151 -> 0.229 | Improved |
| use pi-ID instead of VSD for preference | Qwen-Image Lightning · COCO-10k (Qwen-Image, 4-NFE) | HPSv2.1: 0.322 -> 0.314 | Hurt |
| use pi-ID instead of VSD for text rendering | Qwen-Image Lightning · OneIG-Bench (Qwen-Image, 4-NFE) | Text: 0.923 -> 0.892 | Hurt |
| use data-free instead of data-dependent pi-ID | pi-Flow (GM-Qwen) · COCO-10k (Qwen-Image, 4-NFE) | FID: 36.0 -> 36.0 | No clear effect |
Closest papers by meaning
- Path-Guided Flow Matching for Dataset Distillation (NeurIPS 2026 · similarity 0.87)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.87)
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.87)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.87)
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.87)
- Bespoke Solvers for Generative Flow Models (ICLR 2024 · 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.
