Research map / Image generation / Fast diffusion sampling
Rectified CFG++ for Flow Based Models
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 Rectified-CFG++, a training-free predictor-corrector guidance for rectified-flow text-to-image models that takes a conditional Euler predictor step and then applies a time-scheduled correction using conditional-minus-unconditional velocity evaluated at the intermediate point.
Problem
Standard classifier-free guidance applied to rectified flow (RF) text-to-image models causes off-manifold drift, producing oversaturation, structural distortion, text misalignment and sensitivity to guidance scale.
Approach
Replaces CFG's extrapolation along the guidance difference with a conditional predictor step followed by a scheduled correction alpha(t) times the guidance difference evaluated at the predicted intermediate point, claimed to keep trajectories near the learned transport manifold.
Key results
- On MS-COCO 10K with Flux-dev, Rect-CFG++ lowers FID from 37.86 (CFG) to 32.23 and raises PickScore from 0.3248 to 0.6752.
- On Lumina, FID drops from 26.93 (CFG) to 22.49 with Rect-CFG++ and HPSv2 rises from 0.2797 to 0.3004.
- On MS-COCO 1K with SD3.5, Rect-CFG++ achieves FID 67.15 vs 67.71 for CFG and HPSv2 0.2959 vs 0.2941; gains over CFG are small for SD3.5 and SD3 on several metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CFG with Rect-CFG++ | Lumina · MS-COCO 10K | FID: 26.9321 -> 22.4899 | Helped |
| Replace CFG with Rect-CFG++ | Lumina · MS-COCO 10K | CLIP: 0.3511 -> 0.3464 | Hurt |
| Replace CFG with Rect-CFG++ | SD3 · MS-COCO 10K | FID: 23.8898 -> 23.3945 | Helped |
| Replace CFG with Rect-CFG++ | SD3.5 · MS-COCO 10K | CLIP: 0.3506 -> 0.3497 | Hurt |
| Replace CFG with Rect-CFG++ | Flux-dev · MS-COCO 10K | FID: 37.8625 -> 32.2262 | Helped |
| Replace CFG with Rect-CFG++ | SD3.5 · MS-COCO 1K | FID: 67.7133 -> 67.1495 | Helped |
| Replace CFG with Rect-CFG++ | SD3.5 · MS-COCO 1K | ImageReward: 1.0530 -> 1.0845 | Helped |
| Replace CFG with Rect-CFG++ | SD3.5 · MS-COCO 1K | CLIP: 0.3515 -> 0.3506 | Hurt |
| Replace CFG with Rect-CFG++ | SD3.5 · MS-COCO 1K | HPSv2: 0.2941 -> 0.2959 | Helped |
| Add predictor to the method (ablation) | Rect-CFG++ without predictor · MS-COCO 1K | FID: 73.6981 -> 72.9745 | Helped |
| Add corrector to the method (ablation) | Rect-CFG++ without corrector · MS-COCO 1K | FID: 74.6545 -> 72.9745 | Helped |
| Replace unconditional configuration with Rect-CFG++ (ablation) | Rect-CFG++ · MS-COCO 1K | FID: 91.1180 -> 72.9745 | Helped |
Surprising findings
- On SD3 (MS-COCO 10K), Rect-CFG++ has worse FID than CFG (23.39 vs 23.89 is better, but CLIP and Aesthetic gains are modest) while on SD3.5 FID and CLIP are nearly tied with CFG, suggesting limited gains on some backbones.
- Rect-CFG++ step cost is higher in runtime at 1024x1024 (17.88 s vs 16.26 s for CFG) despite using fewer NFEs (20 vs 28), with FLOPs reported identically.
Limitations
- Theoretical claims rely on strong assumptions (global Lipschitz, bounded guidance and velocity), and the proofs are informal with inconsistent steps.
- No error bars or statistical significance reported.
- Guidance schedule alpha(t) allowed to exceed 1, so the 'interpolation' framing does not strictly hold.
Experimental setup
Datasets: MS-COCO LAION-Aesthetic Pick-A-Pic T2I-CompBench
Benchmarks: MS-COCO 10K MS-COCO 1K T2I-CompBench LAION-Aesthetic 1K Pick-A-Pic
Metrics: FID CLIP score Aesthetic score ImageReward PickScore HPSv2 user preference
Baselines: CFG CFG++ APG CFG-Zero* Conditional flow (unguided)
Builds on: classifier-free guidance rectified flow CFG++ adaptive projected guidance
Closest papers by meaning
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.91)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.90)
- CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.90)
- Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.89)
- Improving the Training of Rectified Flows (NeurIPS 2024 · similarity 0.88)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 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.