Research map / Image generation / Text-to-image diffusion
PC-Flow: Preference Alignment in Flow Matching via Classifier
Shaomeng Wang, He Wang, Longquan Dai, Jinhui Tang
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 |
|---|---|---|---|
| add PC-Flow preference classifier guidance to the base flow model | SD-3.5-M · PartiPrompts | LAION Aesthetics: 5.57 -> 5.94 | Improved |
| add PC-Flow guidance | SD-3.5-M · HPS benchmark | HPS: 27.93 -> 28.95 | Improved |
| add PC-Flow guidance | SD-3.5-M · PartiPrompts | PickScore: 22.53 -> 22.63 | Improved |
| add PC-Flow guidance | SD-3.5-M · HPS benchmark | CLIP score: 33.99 -> 34.45 | Improved |
| compare PC-Flow with Flux1.dev on PickScore | PC-Flow · PartiPrompts | PickScore: 22.80 (Flux1.dev) vs 22.63 (PC-Flow) | Hurt |
| compare PC-Flow with SANA-1.5 on LAION Aesthetics | PC-Flow · PartiPrompts | LAION Aesthetics: 5.96 (SANA-1.5) vs 5.94 (PC-Flow) | No clear effect |
| raise temperature beta from 20 to 200 (ablation) | PC-Flow loss · HPS benchmark | LAION Aesthetics: 5.81 -> 6.14 | Improved |
| use a very small temperature beta of 5 instead of 200 (ablation) | PC-Flow loss · HPS benchmark | PickScore: 22.27 -> 20.34 | Hurt |
| raise temperature beta from 200 to 2000 (ablation) | PC-Flow loss · HPS benchmark | LAION Aesthetics: 6.14 -> 5.87 | Hurt |
| raise guidance weight gamma from 0.05 to 0.10 (ablation) | PC-Flow sampling · HPS benchmark | LAION Aesthetics: 5.84 -> 6.14 | Improved |
| raise guidance weight gamma above 0.10 (to 0.15 and 0.20) (ablation) | PC-Flow sampling · HPS benchmark | LAION Aesthetics: 6.14 -> 5.87 (0.15), 5.85 (0.20) | Hurt |
| use PC-Flow instead of AuraFlow | PC-Flow · HPS benchmark | PickScore win rate (%): 87.08 | Improved |
Closest papers by meaning
- Preference Alignment with Flow Matching (NeurIPS 2024 · similarity 0.93)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.90)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · similarity 0.88)
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.88)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.88)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 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.
