README.md

metadata

license: apache-2.0
pipeline_tag: image-to-image
tags:
  - image-editing
  - multi-modal

Text-Vision Co-Instructed Image Editing (TV-Edit)

This repository contains the weights for TV-Edit, a Textual-Visual instruction unified Editing framework. TV-Edit jointly models textual instructions as semantic intent and sparse visual instructions (such as points or drags) as spatial guidance to achieve precise and intent-faithful image manipulation.

Project Page | Paper | GitHub

Introduction

TV-Edit is designed to unify the strengths of textual instructions and visual prompts. By integrating semantic intent and spatial constraints, TV-Edit leads to more precise spatial control, less instruction ambiguity, and stronger structural consistency than text-only or drag-based alternatives.

Installation

To set up the environment, run the following commands:

## git clone the repository
git clone https://github.com/PolyU-VCLab/TVEdit.git
cd TVEdit

# create an environment
conda create -n TVEdit python=3.10
conda activate TVEdit

pip install --upgrade pip
pip install torch==2.5.0+cu121 torchvision==0.20.0+cu121 --index-url https://download.pytorch.org/whl/cu121
pip install transformers==4.52.4 pytorch-lightning==2.4.0 diffusers==0.35.1

Usage

You can run the interactive Gradio demo to test the model:

1. Download the base model checkpoint: Qwen-Image-Edit.
2. Download the trained TV-Edit weights from this repository.
3. [Optional] TV-Edit supports existing trained acceleration LoRA for 4-step editing: Qwen-Image-Edit-4step.
4. Launch the Gradio demo:

python gradio_demo.py

After launching the demo, you can upload an image, draw point trajectories for spatial control, and enter textual instructions for semantic changes.

Citation

If you find this work helpful, please cite:

@article{xie2026text-vision,
  title={Text-Vision Co-Instructed Image Editing},
  author={Xie, Chenxi and Wu, Yuhui and Yi, Qiaosi and Zhang, Lei},
  journal={arXiv preprint arXiv:2606.16767},
  year={2026},
}