Rectified-CFG++ for Flow Based Models

Authors: Shreshth Saini, Shashank Gupta, Alan C. Bovik
The University of Texas at Austin

Project Webpage: https://shreshthsaini.github.io/Rectified-CFGpp/

A training-free, geometry-aware guidance scheme for flow-based text-to-image (T2I) models. Rectified-CFG++ replaces the naïve extrapolation of classifier-free guidance (CFG) with a predictor–corrector integrator that stays on the learned data manifold, eliminating structural artifacts while improving prompt alignment and over generation quality.

---

✨ Features

• On-Manifold Sampling
  Predictor–corrector updates keep trajectories on the data manifold.
• Training-Free
  Drop-in replacement for standard CFG—no additional training or fine-tuning required.
• Model-Agnostic
  Works with any transformer-based rectified-flow T2I backbone (e.g., Flux, SD3/3.5, Lumina).
• Stable Across Scales
  Maintains visual fidelity and prompt alignment even at high guidance strengths.
• Efficient
  Achieves state-of-the-art FID and CLIP-Score.

---

🛠️ Installation

1. Clone the repository:

git clone https://github.com/shreshthsaini/Rectified-CFGpp.git
cd Rectified-CFGpp

2. Create and activate a virtual environment using the provided YAML configuration file:

conda env create -f environment.yaml
conda activate rectcfgpp

🚀 Quick Start

The repository includes a simple demo script (demo.py) that allows you to quickly test Rectified-CFG++ with various parameters:

Basic Usage

Generate an image with default parameters:

python demo.py --prompt "A magical forest at twilight with glowing fireflies"

Advanced Usage

Customize the generation process with additional parameters:

python demo.py \
    --prompt "A cyberpunk city at night" \
    --guidance_scale 4.5 \
    --num_steps 28 \
    --width 1024 \
    --height 1024 \
    --output_dir "./my_generations" \
    --seed 99

Available Parameters

• --prompt: Text prompt for image generation (required)
• --negative_prompt: Negative prompt for generation (default: "")
• --guidance_scale: Guidance scale for CFG (default: 4.5)
• --num_steps: Number of inference steps (default: 28)
• --sigma_noise: Sigma noise parameter (default: 0.005)
• --width: Output image width (default: 1024)
• --height: Output image height (default: 1024)
• --output_dir: Directory to save generated images (default: "./outputs")
• --seed: Random seed for reproducibility (default: 42)

Generated images will be saved in the specified output directory with filenames indicating the guidance scale and number of steps used.

🖼️ Visual Highlights

Predictor–corrector updates prevent drift away from the learned data manifold.

Representative samples showcasing improved prompt adherence and perceptual quality.

📚 Citation

If you find Rectified-CFG++ useful in your research, please cite our NeurIPS 2025 paper:

@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},
  year      = {2025}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.