Delta Velocity Rectified Flow (DVRF)

Official implementation of the paper: "Delta Velocity Rectified Flow for Text-to-Image Editing".
ArXiv: [https://arxiv.org/abs/2509.05342]

Overview

DVRF is a text-guided image editing method that optimizes the latent of a pre-trained diffusion model (SD3 / SD3.5) using a rectified-flow objective on the delta of predicted velocities between a source and a target prompt. It provides high-fidelity, localized edits while preserving the structure of the source image.

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

---

Repository Structure

DeltaVelocityRectifiedFlow/
├── assets/                     # Paper figures and results
│   ├── DVRF.png               # Method schematic
│   ├── DVRF_results.png       # Qualitative results
│   ├── DVRF_comparaison.png   # Comparison results 1
│   └── DVRF_comparaison2.png  # Comparison results 2
├── images/                     # Example images and dataset config
│   ├── mapping_file.yaml      # Dataset configuration
│   ├── a_cat_sitting_on_a_table.png
│   ├── city-street.jpg
│   ├── fallow-deer.jpg
│   ├── ...
├── models/                     # Core DVRF implementation
│   ├── __init__.py
│   └── DVRF.py                # Main DVRF algorithm
├── edit.py                     # Main script for running experiments
├── exp.yaml                    # Experiment configuration
├── dvrf_environment.yml        # Conda environment
├── .gitignore                  # Git ignore rules
└── README.md                   # This file

---

Method

The DVRF objective optimizes the latent by aligning the target velocity with the source velocity, inspired by the Dela Denoising Score for diffusions models. We further introduce a shift term to improve editing performance, and propose Delta Velocity Rectified Flow (DVRF), a trajectory-driven editing objective that operates in the velocity space of rectified flows. DVRF obtains state-of-the-art results on the PIE Benchmark. See the method schematic:

---

Results

Selected qualitative results demonstrating localized edits and structure preservation:

Additional comparisons:

---

Installation

Clone the repo:

git clone https://github.com/gaspardbd/DeltaVelocityRectifiedFlow.git
cd DeltaVelocityRectifiedFlow

Create the conda environment :

conda env create -f dvrf_environment.yml
conda activate dvrf_env

---

Quick Start

1. Configure your experiment in exp.yaml:

- exp_name: "DVRF_SD3"
  dataset_yaml: images/mapping_file.yaml
  model_type: "SD3"        # or "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             # optimization steps
  seed: 41
  eta: 1.0                  # progressive c_t = k/T * t  described in the paper
  scheduler_strategy: "descending"   # "random" or "descending"
  lr: "custom"             # or a float, e.g. 0.02
  optimizer: "SGD"         # SGD, Adam, AdamW, RMSprop, SGD_Nesterov

2. 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.

3. Run editing:

python edit.py --exp_yaml exp.yaml

Outputs are saved under outputs/<exp_name>/<model_type>/src_<image_name>/tgt_<index>/ including the side-by-side image and trajectory frames.

---

Citation

If you use this code, please cite our paper:

@misc{beaudouin2025deltavelocityrectifiedflow,
      title={Delta Velocity Rectified Flow for Text-to-Image Editing}, 
      author={Gaspard Beaudouin and Minghan Li and Jaeyeon Kim and Sung-Hoon Yoon and Mengyu Wang},
      year={2025},
      eprint={2509.05342},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.05342}, 
}

---

License

http://creativecommons.org/licenses/by/4.0/

---

Acknowledgements

• Built on top of Hugging Face Diffusers pipelines and Stable Diffusion 3/3.5.
• Thanks to the research community for open-source models and tooling.