• [2025.07.30] 🎆 We upload benchmark results and the FluxKontext fine-tune-only T5 checkpoint.
• [2025.07.27] 🎆 We release GPT-Image-Edit, a state-of-the-art image editing model with 1.5M high-quality editing samples. All data, models, training code and evaluation code are open-sourced. Our code is based on UniWorld-V1, Thanks to the authors of UniWorld-V1. Checking our report for more details. Welcome to watch 👀 this repository for the latest updates.

1.Set up environment

git clone https://github.com/wyhlovecpp/GPT-Image-Edit.git
cd GPT-Image-Edit
conda create -n univa python=3.10 -y
conda activate univa
pip install -r requirements.txt
pip install flash_attn --no-build-isolation

2.Download pretrained checkpoint

huggingface-cli download --resume-download UCSC-VLAA/gpt-image-edit-training --local-dir ${MODEL_PATH}
huggingface-cli download --resume-download black-forest-labs/FLUX.1-Kontext-dev --local-dir ${FLUX_PATH}

3.Run with CLI

MODEL_PATH="path/to/model/final_checkpoint"
FLUX_PATH="path/to/flux"
CUDA_VISIBLE_DEVICES=0 python -m univa.serve.cli \
    --model_path ${MODEL_PATH} \
    --flux_path ${FLUX_PATH}

4.Run with Gradio Web Interface Highly recommend trying out our web demo by the following command.

python app.py --model_path ${MODEL_PATH} --flux_path ${FLUX_PATH}

5.Run with Gradio Serve

CUDA_VISIBLE_DEVICES=0 python -m univa.serve.gradio_web_server \
    --model_path ${MODEL_PATH} \
    --flux_path ${FLUX_PATH}

Download the data from UCSC-VLAA/GPT-Image-Edit-1.5M. The dataset consists of two parts: source images and annotation metadata. The json files is under UCSC-VLAA/gpt-image-edit-training/training_json

Prepare a data.txt file in the following format:

1. The first column is the root path to the image.

2. The second column is the corresponding annotation JSON file.

3. The third column indicates whether to enable the region-weighting strategy. We use False in our training setting.

We have prepared a data.txt file about gpt-edit for your reference.

data/gpt-edit/hqedit/edit,training_json/hqedit_gpt_edit.json,false
data/gpt-edit/hqedit/generate,training_json/hqedit_gpt_generate.json,false
data/gpt-edit/omniedit,training_json/omniedit_gpt.json,false
data/gpt-edit/omniedit,training_json/omniedit_gpt_rewrite.json,false
data/gpt-edit/omniedit/complex-edit,training_json/complexedit_gpt.json,false
data/gpt-edit/ultraedit,training_json/ultraedit_gpt.json,false

Download black-forest-labs/FLUX.1-Kontext-dev to $FLUX_PATH. Download Qwen/Qwen2.5-VL-7B-Instruct to $QWENVL_PATH. We also support other sizes of Qwen2.5-VL.

SAVE_PATH="path/to/save/UniWorld-Qwen2.5-VL-7B-Instruct-FLUX.1-dev-fp32"
python scripts/make_univa_qwen2p5vl_weight.py \
    --origin_flux_ckpt_path $FLUX_PATH \
    --origin_qwenvl_ckpt_path $QWENVL_PATH \
    --save_path ${SAVE_PATH}

You need to specify pretrained_lvlm_name_or_path to ${SAVE_PATH} in flux_qwen2p5vl_7b_vlm_stage1_512.yaml.

We recommend using pretrained weight from LanguageBind/UniWorld-V1/stage1.

# stage 1
# if use prodigy, pip install prodigy
bash scripts/denoiser/flux_qwen2p5vl_7b_vlm_stage1_512.sh

You need to specify pretrained_mlp2_path, which is trained by stage 1 or use the pretrained weight from LanguageBind/UniWorld-V1/stage1.

For training with 512×512 scale images (batch size 1), it consume about 78G in 1 node (8 GPUs).

Setting ema_pretrained_lvlm_name_or_path: null can saving memory if you want to train the higher resolution (e.g, 1024×1024 scale) or larger batch size. Using more nodes also can save memory because we use zero2 for main model in stage 2.

# stage 2
bash scripts/denoiser/flux_qwen2p5vl_7b_vlm_stage2_1024.sh

cd univa/eval/imgedit
# follow the instruction in univa/eval/imgedit/README.md

cd univa/eval/gdit
# follow the instruction in univa/eval/gdit/README.md

cd univa/eval/complex-edit
# follow the instruction in univa/eval/complex-edit/README.md

cd univa/eval/omnicontext
# follow the instruction in univa/eval/omnicontext/README.md

• UniWorld-V1: UniWorld-V1 is a unified framework for understanding, generation, and editing.
• ImgEdit: ImgEdit is a large-scale, high-quality image-editing dataset comprising 1.2 million carefully curated edit pairs and a comprehensive benchmark for image editing.
• Complex-edit: Complex-edit is benchmark for complex image editing.
• Qwen2.5-VL: The new flagship vision-language model of Qwen.
• FLUX.1-Kontext-dev: A state-of-the-art image editing model.
• Step1X-Edit: A state-of-the-art image editing model and a comprehensive benchmark for image editing.
• OmniGen2: A state-of-the-art image editing model and a comprehensive benchmark for image editing.

• See LICENSE for details. The FLUX Kontext weights fall under the FLUX.1 Kontext [dev] Non-Commercial License.

@article{wang2025gpt,
  title={GPT-IMAGE-EDIT-1.5 M: A Million-Scale, GPT-Generated Image Dataset},
  author={Wang, Yuhan and Yang, Siwei and Zhao, Bingchen and Zhang, Letian and Liu, Qing and Zhou, Yuyin and Xie, Cihang},
  journal={arXiv preprint arXiv:2507.21033},
  year={2025}
}