Skip to content

README.md

Latest commit

 

History

History
130 lines (90 loc) · 4.2 KB

README.md

File metadata and controls

130 lines (90 loc) · 4.2 KB

Delta Rectified Flow Sampling (DRFS)

Paper | CVPR 2026

Official implementation of "Delta Rectified Flow Sampling for Text-to-Image Editing" (CVPR 2026).

Delta Rectified Flow Sampling for Text-to-Image Editing
Gaspard Beaudouin, Minghan Li, Jaeyeon Kim, Sung-Hoon Yoon, Mengyu Wang

Overview

DRFS is a text-guided image editing method that optimizes the latent representation of a pre-trained rectified-flow model (SD3 / SD3.5) by matching the delta between source and target velocity predictions. Inspired by Delta Denoising Score for diffusion models, DRFS introduces a trajectory-driven editing objective that operates in the velocity space of rectified flows, together with a progressive shift term for improved editing performance. DRFS achieves state-of-the-art results on the PIE Benchmark while preserving the structure of the source image.

  • Models: Stable Diffusion 3 (SD3), Stable Diffusion 3.5 (medium, large, large-turbo)
  • Pipelines: Hugging Face Diffusers
  • Input: Source image + source prompt + target prompt(s)
  • Output: Edited image

Method

DRFS Method

Results

DRFS Results

DRFS Comparison 1

DRFS Comparison 2

Installation

git clone https://github.com/Harvard-AI-and-Robotics-Lab/DeltaRectifiedFlowSampling.git
cd DeltaRectifiedFlowSampling
conda env create -f drfs_environment.yml
conda activate drfs_env

Quick Start

  1. Configure your experiment in exp.yaml:
- exp_name: "DRFS_SD3"
  dataset_yaml: images/mapping_file.yaml
  model_type: "SD3"              # SD3, SD3.5, SD3.5-medium, SD3.5-large, SD3.5-large-turbo
  T_steps: 50                    # diffusion timesteps
  B: 1                           # batch size for averaging the gradient
  src_guidance_scale: 6
  tgt_guidance_scale: 16.5
  num_steps: 50                  # number of optimization steps
  seed: 41
  eta: 1.0                       # progressive c_t = k/T · t (see paper)
  scheduler_strategy: "descending"  # "random" or "descending"
  lr: "custom"                   # adaptive lr from the paper, or a constant float
  optimizer: "SGD"               # SGD, Adam, AdamW, RMSprop, SGD_Nesterov
  1. Prepare images/mapping_file.yaml with your images and prompts:
- input_img: images/a_cat_sitting_on_a_table.png
  source_prompt: A cat sitting on a table.
  target_prompts:
    - A lion sitting on a table.
  1. Run editing:
python edit.py --exp_yaml exp.yaml

Results are saved to outputs/<exp_name>/<model_type>/src_<image>/tgt_<index>/.

Repository Structure

DeltaRectifiedFlowSampling/
├── assets/                     # Paper figures
├── images/                     # Example images and dataset config
│   └── mapping_file.yaml
├── models/
│   ├── __init__.py
│   └── DRFS.py                 # Core DRFS algorithm
├── edit.py                     # Main entry point
├── exp.yaml                    # Experiment configuration
├── drfs_environment.yml        # Conda environment
└── README.md

Citation

If you find this work useful, please cite:

@inproceedings{beaudouin2026drfs,
  title     = {Delta Rectified Flow Sampling for Text-to-Image Editing},
  author    = {Beaudouin, Gaspard and Li, Minghan and Kim, Jaeyeon and Yoon, Sung-Hoon and Wang, Mengyu},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

License

This work is licensed under a Creative Commons Attribution 4.0 International License.

Acknowledgements

Built on Hugging Face Diffusers and Stable Diffusion 3 / 3.5. Thanks to the open-source community.