📋 News

[2025-3-19]: We release the initial version of code and weight, along with documentation and training/inference scripts.

[2025-3-12]: OmniMamba arXiv paper released. Code and Weight are coming soon. Please stay tuned! ☕️

✨ Highlights

• To the best of our knowledge, OmniMamba is the first linear model based unified multimodal understanding and visual generation model.
• OmniMamba achieves competitive performance with only 2M data for training.
• OmniMamba is highly efficient, achieving up to a 119.2 times speedup and 63% GPU memory reduction for long-sequence generation compared to Transformer-based counterparts.

🛠️ Architecture

📊 Qualitative Results

🏁 Getting Started

• Install
• Inference
• Prepare data
• Train

Install

1. Clone this repository and navigate to OmniMamba folder

git clone https://github.com/hustvl/OmniMamba
cd OmniMamba

2. Install Package

# Install PyTorch (with CUDA 11.8) before everything else. those assume you are using cu118
conda create -n omnimamba python=3.10 -y
conda activate omnimamba
pip install torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
wandb login

Inference

Please download our pretrained model at OmniMamba-1.3b

Multimodal Understanding

python scripts/inference_mmu.py --image_path mmu_validation/cat_dog.png --question 'Please describe it in detail.'

Visual Generation

python scripts/inference_t2i.py --prompt 'A bed in a bedroom between two lamps.'

Prepare data

ShareGPT4V: Please refer to the document of the TinyLLaVA to download the dataset without SAM for pretrain.

LVIS-Instruct-4V: Please refer to the document

LRV-Instruct: Please refer to the document

You can download our preprocessed JSON file on Hugging Face.

Folder structure

OmniMamba
├── dataset/
│   ├── pretokenized_coco_train2014.jsonl
│   ├── llava/
│   │   ├── gqa/
│   │   ├── LAION-CC-SBU/
│   │   ├── ocr_vqa/
│   │   ├── POPE/
│   │   ├── share_textvqa/
│   │   ├── textvqa/
│   │   ├── vg/
│   │   ├── web-celebrity/
│   │   ├── web-landmark/
│   │   ├── wikiart/
│   │   ├── coco/
│   │   ├── lrv_Instruct/
│   │   ├── share-captioner_coco_lcs_676k_1121.json
│   │   ├── sharegpt4v_llava_v1_5_lvis4v_lrv_mix1231k.json

Train

Stage 1: MMU Pre-Training

accelerate launch --mixed_precision=bf16 --machine_rank=0 --num_processes=8 --num_machines=1 --main_process_port=8888 train_stage2.py --config config/config_stage1_mmu.yaml

Stage 1: T2I Pre-Training

accelerate launch --mixed_precision=bf16 --machine_rank=0 --num_processes=8 --num_machines=1 --main_process_port=8888 train_stage2.py --config config/config_stage1_t2i.yaml

Stage 2: Unifid Fine-Tuning

accelerate launch --mixed_precision=bf16 --machine_rank=0 --num_processes=8 --num_machines=1 --main_process_port=8888 train_stage2.py --config config/config_stage2.yaml

❤️ Acknowledgements

We build our project based on

• Mamba
• Cobra
• AiM
• Show-o
• LLaVA

Thanks for their great works.

📚 Citation

If you find OmniMamba useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.

@misc{zou2025omnimambaefficientunifiedmultimodal,
      title={OmniMamba: Efficient and Unified Multimodal Understanding and Generation via State Space Models}, 
      author={Jialv Zou and Bencheng Liao and Qian Zhang and Wenyu Liu and Xinggang Wang},
      year={2025},
      eprint={2503.08686},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.08686}, 
}