===================================================================================

本模型是阿里魔搭 https://huggingface.co/modelscope/Nexus-GenV2 官方模型的量化版本，其中 Qwen-VL 部分采用 NF4 量化，微调 generation_decoder 和 edit_decoder 部分采用 float8_e4m3fn 量化，用户可采用官方代码，稍微调整一下模型加载方式的代码，即可在无需量化的情况下直接加载本模型进行推理，大大减少模型下载流量和硬盘占用空间。

This model is a quantized version of the official Ali ModelScope https://huggingface.co/modelscope/Nexus-GenV2. The Qwen-VL part uses NF4 quantization, and the generation_decoder and edit_decoder fine-tuned parts using float8_e4m3fn quantization. Users can through the official code and adjustment a little for model loading method to directly use this model for inference without quantization, which significantly reduces download traffic and disk space usage.

以下引用原官方模型卡信息：

News

• July 11, 2025: Nexus-Gen V2 is released. Please check more details in the technical report. The model is opitimized from the following aspects:
  • Better image understanding capbility (45.7 on MMMU) through optimization on training schedules.
  • Better image generation (0.81 on GenEval) robustness through training with long-short caption.
  • Better reconstruction in image editing tasks. We have proposed a better editing decoder for Nexus-Gen.
  • Support generation and editing with Chinese prompts.
• May 27, 2025: We fine-tuned Nexus-Gen using the BLIP-3o-60k dataset, significantly improving the model's robustness to text prompts in image generation, achieving a GenEval score of 0.79. The model checkpoints have been updated.

What is Nexus-Gen

Nexus-Gen is a unified model that synergizes the language reasoning capabilities of LLMs with the image synthesis power of diffusion models. We propose a unified image embedding spaces to model image understanding, generation and editing tasks. To perform joint optimization across multiple tasks, we curate a large-scale dataset of 26.3 million samples and train Nexus-Gen using a multi-stage strategy, which includes the multi-task pretraining of the autoregressive model and conditional adaptations of the generation and editing decoders.

More information please refer to our repo: https://github.com/modelscope/Nexus-Gen.git

Model Inference

Installation

# 1. Install [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio.git) from source
git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio
pip install -e .

# 2. Install requirements
pip install -r requirements.txt

# 3. Install ms-swift if you want to perform finetuning on Nexus-Gen.
pip install ms-swift==3.3.0.dev0

Prepare models

Nexus-Gen adopts Qwen2.5-VL-Instruct 7B as its autoregressive model, and adopts FLUX.1-Dev as the vision decoders (including the generation decoder and editing decoder). You can run the following scripts to download the checkpoints.

python download_models.py

Image Understanding

Nexus-Gen inheret the image understanding ability of Qwen2.5-VL. Try the following script (Needs at least 17 GB VRAM).

python image_understanding.py --input_image assets/examples/cat.png --instruction "Please give a brief description of the image"

Image Generation

Try the following scripts to perform image generation (Needs at least 24 GB VRAM). Please see image_generation.py for details about the inference hyperparameters.

python image_generation.py --prompt "A cute cat" --width 512 --height 512

Nexus-GenV2 supports generation with chinese prompts. You may further set the Chinese template for image generation by setting --language zh as follows.

python image_generation.py --prompt "一只可爱的猫" --language zh --width 1024 --height 1024

Image Editing

The Nexus-Gen model comprises two decoders: a generation decoder and an editing decoder (recommended). The former directly utilizes the 81-dimensional embeddings output by the autoregressive model to generate images, while the latter additionally incorporates the original image's 324-dimensional embeddings, enabling more accurate reconstruction of unedited regions in the image.

Try the follow script to perform image editing with editing decoder.

python image_editing.py --input_image assets/examples/cat.png  --instruction "Add a pair of sunglasses"

When performing large-region image edits such as conceptual modifications, it is recommended to employ the generation decoder. This approach allows the model's image generation capabilities to directly enhance its editing performance. Try the follow script to perform image editing with generation decoder.

python image_editing.py --input_image assets/examples/cat.png  --instruction "The cat is now running in a forest." --use_generation_decoder

Nexus-Gen also supports image editing using Chinese prompts:

python image_editing.py --input_image assets/examples/cat.png  --instruction "给猫加一副太阳镜"

Please see image_editing.py for details about the inference hyperparameters.

Gradio demo

Try Nexus-Gen with a gradio UI:

python app.py

Model training

We train Nexus-Gen using a multi-stage strategy, which includes the multi-task pretraining of the autoregressive model and conditional adaptations of the generation and editing decoders. The unified message-like dataset format is:

"images": ["xxx.jpg", "xxx.jpg"]
"messages": [
    {"role": "user", "content": "<image> xxx"}, 
    {"role": "assistant", "content": "xxx"},
    {"role": "user", "content": "xxx"},
    {"role": "assistant", "content": "xxx <image>"}
]

See assets/example_datasets for more examples.

1. Multi-task pretraining for autoregressive model

The autoregressive model of Nexus-Gen is trained on image understanding, generation and editing tasks using ms-swift framework. Please refer to assets/example_datasets/llm_dataset.jsonl for the example dataset.

Run the following script to perform finetuning on Nexus-Gen V2. Refer to the script for more configurations.

bash train/scripts/train_autoregressive_model.sh

If you would like to train the autoregressive model from sctrach, just replace the checkpoints of Nexus-Gen V2 with that of Qwen2.5-VL-7B-Instruct. Specially, replace the *.safetensors and models/Nexus-GenV2/model.safetensors.index.json files.

2. Conditional adaptation for generation decoder

Generation decoder is trained by image reconstruction with the 81-token image embeddings. There are two steps to train it.

(1) Prepare for the embedding-image dataset: given the message-like dataset assets/example_datasets/gen_decoder_dataset.jsonl, run the following code to pre-calculate the embeddings for each image and get the embed-like dataset assets/example_datasets/embeds_gen/gen_decoder_embeds_dataset.jsonl

python train/utils/prepare_embeddataset_for_gen.py

(2) Train the generation decoder: run the following script to train generation decoder.

bash train/scripts/train_generation_decoder.sh

Please refer to train/configs/generation_decoder.yaml for detailed configurations.

3. Conditional adaptation for editing decoder

Editing decoder is trained on the ImagePulse dataset. There are two steps to train it. (1) Prepare for the embedding-image dataset: given the message-like dataset assets/example_datasets/edit_decoder_dataset.jsonl, run the following code to pre-calculate the embeddings for the source and target images and get the embed-like dataset assets/example_datasets/embeds_edit/edit_decoder_embeds_dataset.jsonl

PYTHONPATH=$(pwd) python train/utils/prepare_embeddataset_for_edit.py

(2) Train the editing decoder: run the following script to train editing decoder.

bash train/scripts/train_editing_decoder.sh

Please refer to train/configs/editing_decoder.yaml for detailed configurations. Please note that the projector of editing decoder includes a transformer layer, which is initialized from Qwen2.5-VL-7B-Instruct. So it is nessary to download the checkpoints to models/Qwen/Qwen2.5-VL-7B-Instruct:

modelscope download --model Qwen/Qwen2.5-VL-7B-Instruct --local_dir models/Qwen/Qwen2.5-VL-7B-Instruct

Training Datasets

To be published.

Qualitative results of Nexus-Gen

Limitations

• Please note that Nexus-Gen was trained on limited text-to-image data and may not be robust to text prompts.

Citation

@misc{zhang2025nexusgenunifiedmodelimage,
      title={Nexus-Gen: A Unified Model for Image Understanding, Generation, and Editing}, 
      author={Hong Zhang and Zhongjie Duan and Xingjun Wang and Yuze Zhao and Weiyi Lu and Zhipeng Di and Yixuan Xu and Yingda Chen and Yu Zhang},
      year={2025},
      eprint={2504.21356},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.21356v2}, 
}