ARGenSeg: Image Segmentation with Autoregressive Image Generation Model

Xiaolong Wang¹  ·  Lixiang Ru¹  ·  Ziyuan Huang¹  ·  Kaixiang Ji¹
Dandan Zheng¹  ·  Jingdong Chen¹  ·  Jun Zhou¹

¹Ant Group  ·  NeurIPS 2025

🏠 About

We propose a novel AutoRegressive Generation-based paradigm for image Segmentation (ARGenSeg), achieving multimodal understanding and pixel-level perception within a unified framework. Prior works integrating image segmentation into multimodal large language models (MLLMs) typically employ either boundary points representation or dedicated segmentation heads. These methods rely on discrete representations or semantic prompts fed into task-specific decoders, which limits the ability of the MLLM to capture fine-grained visual details. To address these challenges, we introduce a segmentation framework for MLLM based on image generation, which naturally produces dense masks for target objects. We leverage MLLM to output visual tokens and detokenize them into images using an universal VQ-VAE, making the segmentation fully dependent on the pixel-level understanding of the MLLM. To reduce inference latency, we employ a next-scale-prediction strategy to generate required visual tokens in parallel. Extensive experiments demonstrate that our method surpasses prior state-of-the-art approaches on multiple segmentation datasets with a remarkable boost in inference speed, while maintaining strong understanding capabilities. Key Innovations:

• Novel Framework: First segmentation paradigm built on a unified multimodal understanding-generation architecture, eliminating task-specific modules.
• SOTA without Extra Heads: Demonstrates unified MLLMs achieve state-of-the-art segmentation without dedicated segmentation heads.
• Efficiency & Robustness: Proposes next-scale prediction to accelerate inference; reveals coarse-to-fine mask generation inherently enhances robustness.

In this codebase, we release:

• ARGenSeg-8B checkpoint
• Training, evaluation, and inference code

🔥 News

• [2026-05-15] We release the inference code, training code, and checkpoints for ARGenSeg.
• [2025-10-23] We release the paper on arXiv.
• [2025-09-18] ARGenSeg has been accepted by NeurIPS 2025! 🔥🔥🔥

---

📦 Installation

Step 1: Create Conda Environment

conda create -n argenseg python=3.10
conda activate argenseg

Step 2: Install Dependencies

pip install -r requirements.txt

Step 3: Install Flash Attention

wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.7/flash_attn-2.5.7+cu122torch2.2cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
pip install flash_attn-2.5.7+cu122torch2.2cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

Step 4: Download VAR Pretrained Weights

mkdir -p internvl/model/var_vae/pretrained_weights
wget -O internvl/model/var_vae/pretrained_weights/vae_ch160v4096z32.pth \
    https://huggingface.co/FoundationVision/var/resolve/main/vae_ch160v4096z32.pth

Step 5: Download ARGenSeg Checkpoint

Download the checkpoint from HuggingFace and extract it to pretrained/InternVL2_5-ARGenSeg-8B/.

---

🎮 Demo

Referring Expression Segmentation

python demos/seg_demo.py

Segmentation & Chat

python demos/seg_demo_chat.py

---

🏋️ Training

Prepare Training Data

Understanding Data

Follow the InternVL documentation for detailed download instructions.

Example meta_path: example/internvl_1_2_finetune.json

Segmentation Data

Follow the PSALM Dataset Documentation for data preparation.

Example annotation format: example/anns/refcoco.jsonl
Example meta_path: example/data_seg.json

Mixed Training Data

Merge the understanding and segmentation JSON files for mixed training.

Example: example/mix_seg_usd.json

Start Training

sh scripts/train_argenseg.sh

---

📊 Evaluation

Please refer to eval/README.md for detailed evaluation instructions on:

• RefCOCO Series (comprehension & segmentation)
• VQA (TextVQA, VQAv2)
• POPE
• MMMU

---

🔗 Citation

If you find this work useful, please cite:

@article{wang2025argenseg,
  title={ARGenSeg: Image Segmentation with Autoregressive Image Generation Model},
  author={Wang, Xiaolong and Ru, Lixiang and Huang, Ziyuan and Ji, Kaixiang and Zheng, Dandan and Chen, Jingdong and Zhou, Jun},
  journal={arXiv preprint arXiv:2510.20803},
  year={2025}
}

---

👏 Acknowledgements

We sincerely thank the contributors of InternVL, VAR, and PSALM for their foundational work and open-source spirit.

---

📄 License

This project is licensed under the MIT License - see the MIT License file for details.