[ICLR 2026] DragFlow: Unleashing DiT Priors with Region-Based Supervision for Drag Editing

Zihan Zhou^1,*, Shilin Lu^1,*, Shuli Leng¹, Shaocong Zhang¹, Zhuming Lian¹, Xinlei Yu², Adams Wai-Kin Kong¹

¹Nanyang Technological University, ²National University of Singapore

^*Equal Contribution

       

TL;DR

This repo presents DragFlow, the first image drag editor implemented on DiT architecture, which harnesses the strong priors of FLUX for high-quality drag-based editing via a novel region-based paradigm and sets a new SOTA on image drag-editing benchmarks. This repo also provides usage guidance for our proposed dataset, ReD Bench.

Contents

• Update Timeline
• Preparation Process
  • 1. Repo Clone
  • 2. Environment Setup
  • 3. Model Initialization
  • 4. Dataset Preparation
  • 5. Structure Check
• Quick Start
  • a. Demo
  • b. Benchmarking
  • c. Interface Usage (Coming Soon)
• Evaluation Metrics
• Dataset Release
• Acknowledgement
• Citation

📆 Update Timeline

• Oct. 2025 -> Preprint Uploaded
• Jan. 2026 -> Paper Accepted by ICLR 2026 ✨
• Feb. 2026 -> ReD_Bench Dataset Uploaded to HuggingFace
• Feb. 2026 -> DragFlow Framework Full Code Released
• Mar. 2026 -> Interactive Interface & Data Maker Release
• Mar. 2026 -> Full Instruction Accomplish

🛠 Preparation Process

1. Repo Clone

Run the following command to clone the code to your local machine:

git clone --recursive https://github.com/Edennnnnnnnnn/DragFlow.git

2. Environment Setup

Then initialize the environment using Conda:

conda env create -f ./dependencies/dragflow.yaml --yes
conda activate dragflow

The submodule dependencies will be cloned automatically with DragFlow. Run the following command to install:

cd ./dependencies/FireFlow
pip install -e ".[all]"
cd ../..

• Notes: If no submodule folder found, please re-trigger the submodule clone at the repo directory, using

  git submodule update --init --recursive

3. Model Initialization

Then cache the adapter checkpoint using the following command:

huggingface-cli download --resume-download Tencent/InstantCharacter --local-dir ./dependencies/checkpoints --local-dir-use-symlinks False

Other reliant models will be loaded from your local HuggingFace_Hub caches or downloaded automatically during the first run. If needed, you can edit the model registered pathways explicitly in ./framework/config.yaml.

4. Dataset Preparation

• Don't want to cache the entire dataset? Try the Quick Demo instead!

Please download the dataset to the expected location, following the instructions present in Dataset Release.

5. Structure Check

Once the preparation steps above are finished, your folder structure should be similar to the following:

DragFlow
├── assets               
│   │── flowchart.png
│   └── intro.png
├── datasets
│   │── demo
│   └── ReD_Bench
├── dependencies
│   │── dragflow.yaml
│   │── FireFlow
│   └── checkpoints
│       └── instantcharacter_ip-adapter.bin
├── framework
│   │── (9 python scripts)
│   │── config.yaml
│   └── adapter
│       └── (4 python scripts)
└── evaluation
    └── (3 python scripts)

🎯 Quick Start

After finishing the preparation steps and activating the environment, we are ready to move on!

a. Demo

For a quick demo, you can run any of the following commands, with a demo tag (e.g., --demo cat):

python ./framework/bench_dragflow.py --demo cat

• Notes: Sample details can be found in ./datasets/demos. You may want to try different cases, using:

  python ./framework/bench_dragflow.py --demo human
  python ./framework/bench_dragflow.py --demo train
  python ./framework/bench_dragflow.py --demo view
  python ./framework/bench_dragflow.py --demo cartoon

b. Benchmarking

Activate the environment prepared and bench on ReD_Bench. Please make sure that the dataset is located at the expected position, such as ./datasets/ReD_Bench.

conda activate dragflow
python ./framework/bench_dragflow.py

• Notes: The first run may take longer. If the required models are not available locally, they will be cached automatically from Huggingface. You could also state your own dataset pathway by applying --dataset_dir, and use specific GPUs, such as:

  python ./framework/bench_dragflow.py --dataset_dir <path_to_dataset_directory> --device_0 <device_id> --device_1 <device_id>

To broaden accessibility and support diverse deployment scenarios, this release is optimized for consumer-grade GPUs. By leveraging quantization techniques and activation checkpointing, the project now supports compatible execution on configurations using dual 24GB NVIDIA GPUs.

c. Interface Usage

(Coming Soon)

📊 Evaluation Metrics

For evaluation purpose, you can run the following command, where --input_dir indicate the directory path of the dragged images. Please make sure that the dataset is located at the expected position, such as ./datasets/ReD_Bench:

python ./evaluation/evaluation.py --input_dir ./outputs

The outcomes will be auto-generated in a parallel directory concerning the path of your input_dir.

• For example, if your input_dir folder is named ./outputs, the outcome folder will be named ./outputs__eval, a CSV file named eval_scores.csv will be generated inside, which contains all sample outcomes and the mean scores.

You can also apply your own dataset path by setting --dataset_dir, and use specific GPUs, such as:

python ./evaluation/evaluation.py --input_dir <path> --dataset_dir <path> --device_0 <device_id> --device_1 <device_id>

Please note that variations of ±0.002 for IF-related Metrics are considered normal in reproduction, whereas MD-related Metrics may exhibit a variation of ±1 due to randomness caused by the nature of the evaluation methods.

🤗 Dataset Release

Regional-based Dragging (ReD) Bench, consisting of 120 samples annotated with precise drag instructions at both point and region levels. Each manipulation in the dataset is associated with an intention label, selected from relocation, deformation, or rotation.

For quick usage, you can run the following code to acquire the ReD_Bench dataset:

git lfs install
cd ./datasets
git clone https://huggingface.co/datasets/Edennnnn/ReD_Bench
cd ..

Please refer to the HuggingFace page for more details about ReD_Bench usage.

🫡 Acknowledgement

We would like to express our gratitude to the following contributors whose work our code and benchmark are built upon: FLUX, Diffusers, InstantCharacter, and FireFlow.

🔖 Citation

If you find our work useful in your research, please consider citing our work:

@article{zhou2025dragflow,
  title={Dragflow: Unleashing dit priors with region based supervision for drag editing},
  author={Zhou, Zihan and Lu, Shilin and Leng, Shuli and Zhang, Shaocong and Lian, Zhuming and Yu, Xinlei and Kong, Adams Wai-Kin},
  journal={arXiv preprint arXiv:2510.02253},
  year={2025}
}