Accepted: NeurIPS 2025
Rectified-CFG++ is a training-free, geometry-aware guidance scheme for flow-based text-to-image models. By replacing the naïve extrapolation of classifier-free guidance with a predictor–corrector integrator that stays on the learned data manifold, we eliminate structural artifacts while improving prompt alignment, generation quality and sampling efficiency.
Geometric interpretation of our method
Classifier-free guidance (CFG) is the workhorse for steering large diffusion models toward text-conditioned targets, yet its naïve application to rectified-flow (RF) based models provokes severe off-manifold drift, yielding visual artifacts, text misalignment, and brittle behaviour. We present Rectified-CFG++, an adaptive predictor–corrector guidance that couples the deterministic efficiency of rectified flows with a geometry-aware conditioning rule. Each inference step first executes a conditional RF update to stay on the learned transport path, then applies a scheduled interpolation between conditional and unconditional velocity fields. We prove marginal consistency and bounded on-manifold trajectories, ensuring stability across guidance strengths. Extensive experiments on large-scale RF backbones (Flux, Stable Diffusion 3/3.5, Lumina) show that Rectified-CFG++ consistently outperforms standard CFG on FID, CLIP-Score, ImageReward, Aesthetic Score, and HPS-v2—while reducing artifacts and accelerating convergence.
Figure: Visual comparison showing the improvements achieved by Rectified-CFG++ over standard CFG.
Our method introduces a predictor-corrector approach that maintains trajectories on the learned data manifold while providing effective guidance. The key insight is to replace linear extrapolation with scheduled interpolation.
Instead of the naive extrapolation used in standard CFG, which can lead to off-manifold drift, our approach first performs a conditional rectified flow update to stay on the learned transport path, then applies a scheduled interpolation between conditional and unconditional velocity fields. This ensures stable generation while maintaining the benefits of guidance.
Algorithm overview
We conduct comprehensive comparisons between existing guidance methods and our Rectified-CFG++. Standard CFG applies linear extrapolation between conditional and unconditional predictions, which often leads to off-manifold drift in rectified flow models.
Quantitative comparison showing superior performance across multiple metrics
The visual analysis reveals the fundamental difference in guidance behavior. While other methods struggle to generate artifact-free images or poorly align with text prompts, our method maintains natural appearance while improving prompt adherence.
Visual comparison highlighting reduced artifacts and improved quality
Our method produces higher quality images with better text alignment and fewer artifacts compared to standard CFG. We demonstrate improvements across various aspects of visual quality and generation fidelity.
High-quality samples generated using Rectified-CFG++
No Guidance vs. CFG vs. Rectified-CFG++
Quantitative comparison on Flux model
One of the key advantages of Rectified-CFG++ is the significant improvement in text legibility within generated images. Standard CFG often produces blurry, distorted, or illegible text due to off-manifold drift, while our method maintains sharp, readable text by preserving manifold trajectories.
Figure: Comparison of text legibility in generated images. Rectified-CFG++ produces sharper, more readable text compared to standard CFG.
Key Improvements in Text Generation:
We conduct extensive quantitative evaluation across multiple benchmarks and metrics to demonstrate the effectiveness of our approach. Our results show consistent improvements across all evaluation criteria.
Comprehensive evaluation on COCO-10k dataset
Results on T2I-CompBench evaluation
The evaluation spans multiple aspects of generation quality:
Human evaluation demonstrates the preference for our method across different quality aspects.
Figure: User study results showing human preference for Rectified-CFG++.
Analysis of intermediate generation trajectories showing improved stability and convergence.
Figure: Visualization of intermediate generation trajectories comparing CFG vs Rectified-CFG++.
@inproceedings{saini2025rectifiedcfgpp,
title = {Rectified-CFG++ for Flow Based Models},
author = {Shreshth Saini and Shashank Gupta and Alan C. Bovik},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2025}
}
Acknowledgements: Code to be released upon publication. We thank the Flux, Stable Diffusion and Lumina teams for open‐source models and data.