Research map / Image generation / Fast diffusion sampling
Improving Classifier-Free Guidance of Flow Matching via Manifold Projection
Jian-Feng Cai, Haixia Liu, Zhengyi Su, Chao Wang
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 manifold projection (CFG-MP+) on top of flow-matching CFG | DiT-XL-2-256 CFG(D2F) · ImageNet 256, w=2.0, NFE=120 | IS: 71.48 -> 86.26 | Improved |
| add manifold projection to CFG | DiT-XL-2-256 CFG(D2F) · ImageNet 256, w=1.5, NFE=120 | FID: 7.45 -> 7.77 (CFG-MP), 7.91 (CFG-MP+) | No clear effect |
| add manifold projection to CFG | DiT-XL-2-256 CFG(D2F) · ImageNet 256, w=2.5, NFE=60 | FID: 12.85 -> 12.51 (CFG-MP), 12.30 (CFG-MP+) | Improved |
| add manifold projection (CFG-MP/MP+) to CFG | Flux-dev · DrawBench, w=3, NFE=120 | HPSv2: 29.30 -> 30.18 (MP), 30.79 (MP+) | Improved |
| add manifold projection (CFG-MP/MP+) to CFG | SD3.5 · GenEval overall | overall accuracy (%): 58.36 -> 60.81 (MP), 61.41 (MP+) | Improved |
| add manifold projection to CFG | SD3.5 · Pick-a-Pic, ImageReward, small metric gaps (w=3.0, NFE=60) | Pick score: 22.32 -> 22.35 (MP), 22.47 (MP+) | No clear effect |
| use Anderson Acceleration AA(1,1) for the projection fixed-point iteration (ablation) | CFG-MP on DiT-XL-2-256 · prediction-gap relative change | relative change r | Improved |
| use operator G (unconditional step first) instead of H (ablation) | CFG-MP fixed-point iteration · DiT-XL-2-256, K=6 | relative change r: FPI(G) -6% to -5% vs FPI(H) -4% to -2% | Improved |
| increase FPI iterations K per step (ablation) | CFG-MP+ on SD3.5 · HPSv2 / ImageReward | HPSv2 | Improved |
| increase AA window size from 1 to 3 (ablation) | CFG-MP+ with operator G · HPSv2 and IR | HPSv2, ImageReward | Improved |
| vary AA damping factor beta (0.1 to 1) (ablation) | CFG-MP+ · HPSv2 / IR | HPSv2, ImageReward | Mixed |
| vary AA window m and damping beta (m=1, beta=1.0 vs others) (ablation) | CFG-MP+ on DiT-XL-2-256 · relative prediction-gap reduction, K=3 | -r (%): 9.31 (m=1, beta=0.2) -> 9.44 (m=1, beta=1.0) | No clear effect |
Closest papers by meaning
- Rectified CFG++ for Flow Based Models (NeurIPS 2025 · similarity 0.93)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.92)
- Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.90)
- CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.90)
- On the Guidance of Flow Matching (ICML 2025 · similarity 0.90)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.89)
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.
