LLaDA-o

LLaDA-o

We introduce LLaDA-o, an effective and length-adaptive omni diffusion model for unified multimodal understanding and generation.

LLaDA-o extends diffusion language modeling to a broader multimodal setting, supporting both visual understanding and visual generation within a single framework. The released codebase provides a practical inference pipeline for interleaved text-image processing and a notebook-based workflow for reproducible experiments.

It was presented in the paper LLaDA-o: An Effective and Length-Adaptive Omni Diffusion Model.

Code: https://github.com/ML-GSAI/LLaDA-o

Highlights

• Unified multimodal modeling for both understanding and generation
• Support for text-to-image generation
• Support for image understanding
• Support for instruction-based image editing
• Reproducible inference workflow through multimodal_demo.ipynb

Supported Tasks

The current release is designed for the following multimodal inference settings:

• Text-to-image: generate images from natural language prompts
• Image understanding: produce textual responses conditioned on an input image
• Image editing: edit an image according to a textual instruction
• Interleaved multimodal inference: process text and image context within a shared diffusion-based framework

Quick Start

Please first download the model checkpoint locally, then use the official repository for inference:

git clone https://github.com/ML-GSAI/LLaDA-o
cd LLaDA-o
bash init_env.sh 

The recommended inference entry point is:

• multimodal_demo.ipynb

In the notebook, set:

MODEL_PATH = "/path/to/local/GSAI-ML-LLaDA-o"

and run the cells sequentially to perform text-to-image generation, image understanding, and image editing.

Notes

• The current inference pipeline expects a local checkpoint path.
• The released demo is intended for GPU-based inference.
• For a complete inference workflow and implementation details, please refer to the official GitHub repository.

Citation

If you find LLaDA-o useful in your research, please consider citing:

@article{you2026lladao,
  title={LLaDA-o: An Effective and Length-Adaptive Omni Diffusion Model},
  author={You, Zebin and Zhang, Xiaolu and Zhou, Jun and Li, Chongxuan and Wen, Ji-Rong},
  journal={arXiv preprint arXiv:2603.01068},
  year={2026}
}

Contact

If you have any questions, please feel free to contact us at [email protected].